maxmustermann/NeMo
1
0
Fork 0
Code Issues Pull requests Projects Releases Wiki Activity

Updates in IR+QA example

Browse source 
Signed-off-by: Vlad Getselevich <vgetselevich@nvidia.com>
This commit is contained in:
Vlad Getselevich 2021-01-13 12:13:56 -08:00
parent c1afd235c8
commit dab33d8a17
10 changed files with 877 additions and 126 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

101
examples/nlp/conditional_generation/abstractive_qa.py
View file

@ -15,14 +15,8 @@
from lfqa_utils import *
def full_test():
# eli5 = nlp.load_dataset('eli5')
# wiki40b_snippets = nlp.load_dataset('wiki_snippets', name='wiki40b_en_100_0')['train']
marvel_snippets = load_data("/home/vgetselevich/data/marvel/wiki/", 100)
# print(eli5['test_eli5'][12345])
# print(wiki40b_snippets[8991855])
def full_test(directory, proj_name, questions, ir_results=5, use_eqa=True):
content_snippets = load_data(directory, 100)
# load models
# dense IR model
@ -31,30 +25,30 @@ def full_test():
_ = qar_model.eval()
# prepare IR index
if not os.path.isfile('marvel_passages_reps_32_l-8_h-768_b-512-512.dat'):
if not os.path.isfile(f'{proj_name}_passages_reps_32_l-8_h-768_b-512-512.dat'):
print("*** Generating dense index ***")
make_qa_dense_index_text_chunks(
qar_model,
qar_tokenizer,
# wiki40b_snippets,
marvel_snippets,
content_snippets,
device='cuda:0',
index_name='marvel_passages_reps_32_l-8_h-768_b-512-512.dat',
index_name=f'{proj_name}_passages_reps_32_l-8_h-768_b-512-512.dat',
)
# load index to memory
faiss_res = faiss.StandardGpuResources()
marvel_passage_reps = np.memmap(
'marvel_passages_reps_32_l-8_h-768_b-512-512.dat',
content_passage_reps = np.memmap(
f'{proj_name}_passages_reps_32_l-8_h-768_b-512-512.dat',
dtype='float32',
mode='r',
# shape=(wiki40b_snippets.num_rows, 128),
shape=(len(marvel_snippets), 128),
shape=(len(content_snippets), 128),
)
marvel_index_flat = faiss.IndexFlatIP(128)
marvel_gpu_index = faiss.index_cpu_to_gpu(faiss_res, 1, marvel_index_flat)
marvel_gpu_index.add(marvel_passage_reps)
content_index_flat = faiss.IndexFlatIP(128)
content_gpu_index = faiss.index_cpu_to_gpu(faiss_res, 1, content_index_flat)
content_gpu_index.add(content_passage_reps)
# generative model
qa_s2s_tokenizer = AutoTokenizer.from_pretrained('yjernite/bart_eli5')
@ -62,17 +56,12 @@ def full_test():
_ = qa_s2s_model.eval()
# run examples
questions = [
"who is iron man?",
"who is Tony Stark?",
"who directed iron man?",
]
answers = []
for question in questions:
# create support document with the dense index
doc, res_list = query_qa_dense_index(
question, qar_model, qar_tokenizer, marvel_snippets, marvel_gpu_index, n_results=5, device='cuda:0'
question, qar_model, qar_tokenizer, content_snippets, content_gpu_index, n_results=ir_results, device='cuda:0'
)
# concatenate question and support document into BART input
question_doc = "question: {} context: {}".format(question, doc)
@ -91,9 +80,11 @@ def full_test():
print(question)
print("Generative Answer: " + gen_answer)
print("Context: " + doc.replace("<P> ", ""))
ex_answer_span, ex_answer_sent = extractive_qa(question, doc.replace("<P> ", ""))
print("Extractive Answer: " + ex_answer_span + " - " + ex_answer_sent)
# print("Context: " + doc.replace("<P> ", ""))
if use_eqa:
ex_answer_span, ex_answer_context = extractive_qa(question, doc.replace("<P> ", ""))
print("Extractive Answer: " + ex_answer_span)
def short_test():
@ -151,48 +142,19 @@ def short_test():
print(answers)
def extractive_qa(question, context):
'''
Uses Extractive QA server.
Input: question and context
Response: short span with the answer and full sentence that contains this span or 'Not found'
'''
url = "http://35.193.226.52:9000/longformqa"
payload = (
f'{{\r\n"question": "{question}",\r\n\"context\": "{context}",\r\n'
f'"maxLength": 128,\r\n"minLength": 1,\r\n "numBeams": 8\r\n}}'
)
headers = {
'Content-Type': "application/json",
'User-Agent': "PostmanRuntime/7.18.0",
'Accept': "*/*",
'Cache-Control': "no-cache",
'Postman-Token': "30549445-ea95-416a-bd74-8ab14ad2036d,0dd97345-55e6-4517-970c-f7032bd1a0d3",
'Host': "35.193.226.52:9000",
'Accept-Encoding': "gzip, deflate",
'Content-Length': "1513",
'Connection': "keep-alive",
'cache-control': "no-cache",
}
print(payload)
response = requests.request("POST", url, data=payload, headers=headers)
if response.ok == True:
print(response.text)
response_dict = json.loads(response.text)
return (response_dict['result_eqa'], response_dict['context'])
else:
print("Answer not found in the context")
return ('Not found', 'Not found')
if __name__ == '__main__':
# load_data("/home/vgetselevich/data/marvel/wiki/", 100)
# short_test()
full_test()
questions = [
"what is minecraft?",
"who created minecraft game?",
"which modes minecraft game has?",
]
# question = "Who were the survivors of Titan?"
question = "Who is Vlad?"
# load_data("/home/vgetselevich/data/minecraft/", 100)
# short_test()
full_test("/home/vgetselevich/data/minecraft/", "minecraft", questions, ir_results=1, use_eqa=True)
'''
question = "Who is Stark?"
context = (
"Anthony Edward Stark is a fictional character portrayed by Robert Downey Jr in the Marvel Cinematic "
"Universe (MCU) film franchise based on the Marvel Comics character of the same name commonly known "
@ -209,4 +171,9 @@ if __name__ == '__main__':
"and his armies, who traveled through time to collect the Infinity Stones, saving the universe "
"from decimation, and leaving behind a legacy as one of Earth's most revered superheroes."
)
# extractive_qa(question, context)
answer_span, context = extractive_qa(question, context)
print(answer_span)
full_answer, content = answer_extender(question, answer_span)
print(full_answer)
'''

113
examples/nlp/conditional_generation/data/iron_man/iron_man_1.txt Normal file
View file

@ -0,0 +1,113 @@
Iron Man is a 2008 American superhero film based on the Marvel Comics character of the same name. Produced by Marvel Studios and distributed
by Paramount Pictures. It is the first film in the Marvel Cinematic Universe. It was directed by Jon Favreau from a screenplay
by the writing teams of Mark Fergus and Hawk Ostby, and Art Marcum and Matt Holloway, and stars Robert Downey Jr. as Tony Stark / Iron Man
alongside Terrence Howard, Jeff Bridges, Shaun Toub, and Gwyneth Paltrow. In the film, following his escape from captivity by a terrorist group,
world famous industrialist and master engineer Tony Stark builds a mechanized suit of armor and becomes the superhero Iron Man.
A film featuring the character was in development at Universal Pictures, 20th Century Fox, and New Line Cinema at various times since 1990,
before Marvel Studios reacquired the rights in 2006. Marvel put the project in production as its first self-financed film,
with Paramount Pictures distributing. Favreau signed on as director in April 2006, and faced opposition from Marvel when trying
to cast Downey in the title role; the actor was signed in September. Filming took place from March to June 2007, primarily in California
to differentiate the film from numerous other superhero stories that are set in New York City-esque environments.
During filming, the actors were free to create their own dialogue because pre-production was focused on the story and action.
Rubber and metal versions of the armor, created by Stan Winston's company, were mixed with computer-generated imagery to create the title character.
Iron Man premiered in Sydney on April 14, 2008, and was released in the United States on May 2, as the first film in Phase One of the MCU.
It grossed over $585 million on its $140 million budget, becoming the eighth-highest grossing film of 2008.
The film was praised by critics for its acting (particularly Downey's), screenplay, direction, visual effects, and action sequences.
It was selected by the American Film Institute as one of the ten best films of 2008 and received two nominations at the 81st Academy Awards
for Best Sound Editing and Best Visual Effects. It was followed by two sequels, Iron Man 2 in 2010 and Iron Man 3 in 2013.
Plot
Tony Stark, who has inherited the defense contractor Stark Industries from his father Howard Stark, is in war-torn Afghanistan with his friend
and military liaison, Lieutenant Colonel James Rhodes, to demonstrate the new "Jericho" missile.
After the demonstration, the convoy is ambushed and Stark is critically wounded by a missile used by the attackers: one of his company's own.
He is captured and imprisoned in a cave by a terrorist group called the Ten Rings. Yinsen, a fellow captive doctor, implants an electromagnet
into Stark's chest to keep the shrapnel shards that wounded him from reaching his heart and killing him. Ten Rings leader Raza offers Stark
freedom in exchange for building a Jericho missile for the group, but he and Yinsen know that Raza will not keep his word.
Stark and Yinsen secretly build a small, powerful electric generator called an arc reactor to power Stark's electromagnet and a prototype suit
of powered armor to aid in their escape. Although they keep the suit hidden almost to completion, the Ten Rings discover their hostages'
intentions and attack the workshop. Yinsen sacrifices himself to divert them while the suit fully charges. The armored Stark battles his way
out of the cave to find the dying Yinsen, then burns the Ten Rings' weapons in anger and flies away, crashing in the desert and destroying the suit.
After being rescued by Rhodes, Stark returns home and announces that his company will cease manufacturing weapons. Obadiah Stane,
his father's old partner and the company's manager, advises Stark that this may ruin Stark Industries and his father's legacy.
In his home workshop, Stark builds a sleeker, more powerful version of his improvised armor suit as well as a more powerful arc reactor for it
and his chest. Personal assistant Pepper Potts places the original reactor inside a small glass showcase. Though Stane requests details,
a suspicious Stark decides to keep his work to himself.
At a charity event held by Stark Industries, reporter Christine Everhart informs Stark that his company's weapons were recently delivered
to the Ten Rings and are being used to attack Yinsen's home village, Gulmira. Stark dons his new armor and flies to Afghanistan,
where he saves the villagers. While flying home, Stark is attacked by two fighter jets. He reveals his secret identity to Rhodes over the phone
in an attempt to end the attack. Meanwhile, the Ten Rings gather the pieces of Stark's prototype suit and meet with Stane,
who has been trafficking arms to criminals worldwide and has staged a coup to replace Stark as Stark Industries' CEO by hiring the Ten Rings
to kill him. He subdues Raza and has the rest of the group killed. Stane has a massive new suit reverse engineered from the wreckage.
Seeking to track his company's illegal shipments, Stark sends Potts to hack into its database. She discovers that Stane hired the Ten Rings
to kill Stark, but the group reneged when they realized they had a direct route to Stark's weapons. Potts meets with Agent Phil Coulson
of S.H.I.E.L.D., an intelligence agency, to inform him of Stane's activities.
Stane's scientists cannot duplicate Stark's miniaturized arc reactor, so Stane ambushes Stark at his home and takes the one from his chest.
Stark manages to get to his original reactor to replace it. Potts and several S.H.I.E.L.D. agents attempt to arrest Stane, but he dons his suit
and attacks them. Stark fights Stane but is outmatched without his new reactor to run his suit at full capacity. The fight carries Stark
and Stane to the top of the Stark Industries building, and Stark instructs Potts to overload the large arc reactor powering the building.
This unleashes a massive electrical surge that causes Stane and his armor to fall into the exploding reactor, killing him. The next day,
at a press conference, Stark defies suggestions from S.H.I.E.L.D. and publicly admits to being "Iron Man."
In a post-credits scene, S.H.I.E.L.D. Director Nick Fury visits Stark at home, telling him that Iron Man is not "the only superhero in the world",
and explaining that he wants to discuss the "Avenger Initiative".
Cast
Robert Downey Jr. as Tony Stark / Iron Man:
An industrialist, genius inventor, and consummate playboy, he is CEO of Stark Industries and chief weapons manufacturer for the U.S. military.
Director Jon Favreau felt that Downey's past made him an appropriate choice for the part and that the actor could not only make Stark
a "likable asshole," but also portray an authentic emotional journey, once he had won over the audience. Favreau was also attracted
to Downey because of his performance in Kiss Kiss Bang Bang. Downey frequently spoke with that film's director, Shane Black,
about the script and dialogue in Iron Man. Downey had an office next to Favreau during pre-production, which allowed him greater
involvement in the screenwriting process, specially when it came to adding humor to the film.
Downey explained, "What I usually hate about these superhero movies is when suddenly the guy that you were digging turns into Dudley Do-Right,
and then you're supposed to buy into all his 'Let's go do some good!' That Eliot Ness-in-a-cape-type thing. What was really important to me
was to not have him change so much that he's unrecognizable. When someone used to be a schmuck and they're not anymore,
hopefully they still have a sense of humor." To get into shape, Downey spent five days a week weight training and practiced martial arts,
which he said benefited him because "it's hard not to have a personality meltdown after about several hours in that suit.
I'm calling up every therapeutic moment I can think of to just get through the day."
Terrence Howard as James "Rhodey" Rhodes:
A friend of Stark's and the liaison between Stark Industries and the United States Air Force in the department of acquisitions,
specifically weapons development. Favreau cast Howard because he felt he could play War Machine in a sequel.
Howard prepared for the role by visiting Nellis Air Force Base on March 16, 2007, where he ate with the pilots and observed HH-60 Pave Hawk
rescue helicopters and F-22 Raptors. While Rhodes is roguish in the comics after he meets Stark, his previous role as a disciplinarian
creates a dynamic tension with Stark's character. He is unsure whether Stark's actions are acceptable.
"Rhodey is completely disgusted with the way Tony has lived his life, but at a certain point he realizes that perhaps there is a different way,
" Howard said. "Whose life is the right way: Is it the strict military life, or the life of an independent?"
Howard and his father are Iron Man fans, partly because Rhodes was one of the few black superheroes when Howard was a child.
He has been a Downey fan since he saw him in Weird Science; the two competed physically on set.
Jeff Bridges as Obadiah Stane / Iron Monger:
Stark's business second-in-command, mentor, and friend, who turns on him to take over the company, eventually building a giant exosuit
to fight Stark. Bridges read the comics as a boy and liked Favreau's modern, realistic approach. He shaved his head, something he had wanted
to do for some time, and grew a beard for the role. Bridges googled the Book of Obadiah, and was surprised to learn retribution is a major theme
in that book of the Bible, something that Stane represents. Many of Stane's scenes were cut to focus more on Stark, but the writers
felt Bridges's performance allowed the application of "less is more" when editing the film.
Shaun Toub as Ho Yinsen:
Stark's fellow captive, who grafts an electromagnet to Stark's chest "to keep the shrapnel shell shards that wounded him from reaching
his heart and killing him" and helps Stark build the first Iron Man suit.
Gwyneth Paltrow as Virginia "Pepper" Potts:
Stark's personal assistant and budding love interest. Paltrow asked Marvel to send her any comics they would consider relevant to her
understanding of the character, whom she considered to be very smart, levelheaded, and grounded. She said she liked "the fact that
there's a sexuality that's not blatant.” Favreau wanted Potts' and Stark's relationship to be reminiscent of a 1940s comedy,
something which Paltrow considered to be fun in an "innocent yet sexy" way.
Additionally, Faran Tahir appears as Raza, the leader of the Ten Rings;
Paul Bettany voices J.A.R.V.I.S., Stark's personal AI system;
Leslie Bibb portrays Christine Everhart, a reporter for Vanity Fair;
and Clark Gregg appears as Phil Coulson, an agent of S.H.I.E.L.D.
Will Lyman provides the voice-over during the opening award ceremony.
Director Jon Favreau plays Happy Hogan, Stark's bodyguard and chauffeur, and Samuel L. Jackson makes a cameo appearance as Nick Fury,
director of S.H.I.E.L.D., in a post-credits scene. Jackson's face was previously used as the model for the Ultimate Marvel imprint version
of Nick Fury.
Other cameos in the film include Stan Lee as himself, being mistaken for Hugh Hefner by Stark at a party;
Peter Billingsley as William Ginter Riva, a scientist who works for Stane;
Tom Morello, who provided guitar music for the film, as a terrorist guard and Jim Cramer as himself.
Ghostface Killah had a cameo in a scene where Stark briefly stays in Dubai, but the scene was cut from the theatrical release for pacing reasons.

121
examples/nlp/conditional_generation/data/iron_man/iron_man_2.txt Normal file
View file

@ -0,0 +1,121 @@
Iron Man 2 is a 2010 American superhero film based on the Marvel Comics character Iron Man. Produced by Marvel Studios and distributed
by Paramount Pictures. It is the sequel to Iron Man (2008) and the third film in the Marvel Cinematic Universe (MCU).
Directed by Jon Favreau and written by Justin Theroux, the film stars Robert Downey Jr. as Tony Stark / Iron Man alongside Gwyneth Paltrow,
Don Cheadle, Scarlett Johansson, Sam Rockwell, Mickey Rourke, and Samuel L. Jackson. Six months after Iron Man, Tony Stark resists calls
from the United States government to hand over the Iron Man technology, which is causing his declining health. Meanwhile,
Russian scientist Ivan Vanko uses his own version of the technology to pursue a vendetta against the Stark family.
Following the critical and commercial success of Iron Man in May 2008, Marvel Studios announced and immediately set to work on producing a sequel.
In July, Theroux was hired to write the script and Favreau was signed to return as director. Downey, Paltrow, and Jackson were set to reprise
their roles from Iron Man, while Cheadle was brought in to replace Terrence Howard in the role of James Rhodes. In the early months of 2009,
Rourke (Vanko), Rockwell, and Johansson filled out the supporting cast. Filming took place from April to July 2009, mostly in California
like the first film except for a key sequence in Monaco. Unlike its predecessor, which mixed digital and practical effects,
the sequel primarily relied on computer-generated imagery to create the Iron Man suits.
Iron Man 2 premiered at the El Capitan Theatre on April 26, 2010, and was released in the United States on May 7, as part of Phase One of the MCU.
The film received generally positive reviews for Downey's performance as well as the action sequences, musical score, and visual effects,
but many critics deemed it to be inferior to the first film and criticized the villain and pacing.
The sequel grossed over $623.9 million at the worldwide box office, making it the seventh-highest-grossing film of 2010.
It received an Academy Award nomination for Best Visual Effects. A third installment of the series, Iron Man 3, was released on May 3, 2013.
Plot
In Russia, the media covers Tony Stark's disclosure of his identity as Iron Man. Ivan Vanko, whose father, Anton Vanko, has just died,
sees this and begins building a miniature arc reactor similar to Stark's. Six months later, Stark is a superstar and uses his Iron Man suit
for peaceful means, resisting government pressure to sell his designs. He reinstitutes the Stark Expo in Flushing Meadows, New York to continue
his father Howard's legacy.
Later, Stark learns that the palladium core in the arc reactor that keeps him alive and powers the armor is slowly poisoning him,
and he cannot find a substitute. Growing increasingly reckless and despondent about his impending death, and choosing not to tell anyone
about his condition, Stark appoints his personal assistant Pepper Potts as CEO of Stark Industries, and hires Stark employee Natalie Rushman
to replace her as his personal assistant. Stark competes in the Monaco Historic Grand Prix, where he is attacked in the middle of the race by Vanko,
wielding electrified whips. Stark dons his armor and defeats Vanko, but the suit is severely damaged. Vanko explains his intention was to prove
to the world that Iron Man is not invincible. Impressed by Vanko's performance, Stark's rival, Justin Hammer, fakes Vanko's death while
breaking him out of prison and asks him to build a line of armored suits to upstage Stark. During what he believes is his final birthday party,
Stark gets drunk while wearing the Iron Man suit. Disgusted, Stark's best friend, U.S. Air Force Lieutenant Colonel James Rhodes,
dons Stark's prototype armor and tries to restrain him. The fight ends in a stalemate, and Rhodes confiscates the armor for the U.S. Air Force.
Nick Fury, director of S.H.I.E.L.D., approaches Stark, revealing that "Rushman" is Agent Natasha Romanoff and that Howard Stark was
a S.H.I.E.L.D. founder whom Fury knew personally. Fury explains that Vanko's father and Stark invented the arc reactor together, but when
Anton tried to sell it, Stark had him deported. The Soviets sent Anton to the Gulag. Fury gives Stark some of his father's old material.
Tony discovers a hidden message in the diorama of the 1974 Stark Expo; it proves to be a diagram of the atomic structure of a new element.
With the aid of his A.I., J.A.R.V.I.S., Stark synthesizes it. When he learns that Vanko is still alive, he places the new element in his arc
reactor and ends his palladium dependency. At the Expo, Hammer unveils Vanko's armored drones, led by Rhodes in a heavily weaponized version
of the prototype armor. Stark arrives to warn Rhodes, but Vanko takes remote control of all the drones and Rhodes' armor and attacks Stark.
Hammer is arrested while Romanoff and Stark's bodyguard Happy Hogan go after Vanko at Hammer's factory. Vanko escapes, but Romanoff returns
control of Rhodes' armor to him. Together, Stark and Rhodes defeat Vanko and his drones. Vanko commits suicide by blowing up his suit,
along with the defeated drones. At a debriefing, Fury informs Stark that because of his difficult personality, S.H.I.E.L.D. intends
to use him only as a consultant. Stark and Rhodes receive medals for their heroism. In a post-credits scene, S.H.I.E.L.D. agent Phil Coulson
reports the discovery of a large hammer at the bottom of a crater in a desert in New Mexico.
Cast
Robert Downey Jr. as Tony Stark / Iron Man:
A billionaire who escaped captivity in Afghanistan with a suit of armor he created, he now struggles to keep his technology out of
the government's hands. Downey and Favreau, who had been handed a script and worked from it on the first movie, conceived part of
the film's story themselves. On Stark being a hero, Downey said "It's kind of heroic, but really kind of on his own behalf. So I think
there's probably a bit of an imposter complex and no sooner has he said, 'I am Iron Man–' that he's now really wondering what that means.
If you have all this cushion like he does and the public is on your side and you have immense wealth and power, I think he's way too insulated
to be okay." Downey put on 20 pounds of muscle to reprise the role. Six-year-old Davin Ransom portrays Tony Stark as a child.
Gwyneth Paltrow as Pepper Potts:
Stark's closest friend, budding love interest, and business partner; Pepper is promoted to CEO of Stark Industries. On her character's promotion,
Paltrow opined "When we start Iron Man 2 Pepper and Tony are very much in the same vibe as the movie progresses,
Pepper is given more responsibility and she's promoted and it's nice to see her sort of grow up in that way.
I think it really suits her, the job fits her really well."[13] Paltrow expressed excitement about working with Johansson.
Don Cheadle as James "Rhodey" Rhodes / War Machine:
An officer in the U.S. Air Force and Tony Stark's close personal friend who later operates the War Machine armor.
Cheadle replaces Terrence Howard from the first film. Cheadle only had a few hours to accept the role and did not even know what
storyline Rhodes would undergo. He commented that he is a comic book fan, but had not previously participated in comics-themed
films due to the scarcity of black superheroes. Cheadle said he thought Iron Man was a robot before the first film came out.
On how he approached his character, Cheadle stated "I go, what's the common denominator here? And the common denominator was really
his friendship with Tony, and that's what we really tried to track in this one. How is their friendship impacted once Tony comes out
and owns 'I am Iron Man'?". Cheadle said his suit was 23 kilograms (50 lb) of metal, and that he could not touch his face while wearing it.
Cheadle signed a six-picture deal.
Scarlett Johansson as Natasha Romanoff / Black Widow:
An undercover spy for S.H.I.E.L.D. posing as Stark's new assistant. Johansson dyed her hair red before she landed the part, hoping that
it would help convince Favreau that she was right for the role. On why she chose the role, Johansson said, "the Black Widow character
resonated with me. She is a superhero, but she's also human. She's small, but she's strong. She is dark and has faced death so many times
that she has a deep perspective on the value of life. It's hard not to admire her." She stated that she had "a bit of a freak-out moment"
when she first saw the cat-suit. When asked about fighting in the costume, Johansson responded "a big part of me is like 'can I move in this?
Can I run in it? Can I like throw myself over things with this?' And I think just the prep, you just have to put in the hours.
That's what I realized is that just putting in the hours and doing the training and repetition and basically just befriending the stunt team
and spending all day, every day, just over and over and over and over until you sell it."
Sam Rockwell as Justin Hammer:
A rival weapons manufacturer. Sam Rockwell was considered for the role of Tony Stark in the first film, and he accepted the role of Hammer
without reading the script. He had never heard of the character before he was contacted about the part, and was unaware Hammer is
an old Englishman in the comics. Rockwell said, "I worked with Jon Favreau on this film called Made. And Justin Theroux, who wrote the script,
is an old friend of mine, they sort of cooked up this idea and pitched it to Kevin Feige. What they did, they were maybe going to do one villain
like they did with Jeff Bridges, but then they decided to split the villains. And really Mickey Rourke is the main villain, but I come to his aid.
Rockwell described his character as "plucky comic relief, but he's got a little bit of an edge".
Mickey Rourke as Ivan Vanko / Whiplash:
A Russian physicist and ex-convict who builds a pair of arc reactor-based electric whips to exact vengeance on the Stark family.
The character is an amalgam of Whiplash and Crimson Dynamo. Rourke visited Butyrka prison to research the role, and he suggested half
of the character's dialogue be in Russian. He also suggested the addition of tattoos, gold teeth and a fondness for a pet cockatoo,
paying for the teeth and bird with his own money. Rourke explained that he did not want to play a "one-dimensional bad guy", and wanted
to challenge the audience to see something redeemable in him. Not knowing anything about computers, Rourke described pretending to be
tech-savvy as the hardest part of the role.
Samuel L. Jackson as Nick Fury:
Director of S.H.I.E.L.D.; Jackson signed a nine-film contract to play the character. On the subject of his character not seeing any action
in the film, Jackson said "We still haven't moved Nick Fury into the bad-ass zone. He's still just kind of a talker."
The director, Jon Favreau, reprises his role as Happy Hogan, Tony Stark's bodyguard and chauffeur, while Clark Gregg and Leslie Bibb
reprise their roles as S.H.I.E.L.D. Agent Phil Coulson and reporter Christine Everhart, respectively.
John Slattery appears as Tony's father Howard Stark and Garry Shandling appears as United States Senator Stern,
who wants Stark to give Iron Man's armor to the government. Favreau stated that Shandling's character was named after radio
personality Howard Stern. Paul Bettany again voices Stark's computer, J.A.R.V.I.S. Olivia Munn originally appeared as an unnamed character
who was subsequently cut from the film.
Favreau then gave her the role of Chess Roberts, a reporter covering the Stark expo.
Yevgeni Lazarev appears as Ivan Vanko's father Anton Vanko, Kate Mara portrays a Process Server who summons Tony to the government hearing,
and Stan Lee appears as himself (but is mistaken for Larry King).
Additionally, news anchor Christiane Amanpour and political commentator Bill O'Reilly play themselves in newscasts.
Adam Goldstein appears as himself and the film is dedicated to his memory.
Further cameos include Tesla Motors CEO Elon Musk and Oracle Corporation CEO Larry Ellison.
Favreau's son Max appears as a child wearing an Iron Man mask whom Stark saves from a drone.
This was retroactively made as the introduction of a young Peter Parker to the MCU, as confirmed in June 2017 by eventual Spider-Man actor
Tom Holland, Feige and Spider-Man: Homecoming director Jon Watts.

165
examples/nlp/conditional_generation/data/iron_man/iron_man_3.txt Normal file
View file

@ -0,0 +1,165 @@
Iron Man 3 is a 2013 American superhero film based on the Marvel Comics character Iron Man, produced by Marvel Studios and distributed
by Walt Disney Studios Motion Pictures. It is the sequel to Iron Man (2008) and Iron Man 2 (2010), and the seventh film in the Marvel
Cinematic Universe. The film was directed by Shane Black from a screenplay he co-wrote with Drew Pearce, and stars Robert Downey Jr.
as Tony Stark / Iron Man alongside Gwyneth Paltrow, Don Cheadle, Guy Pearce, Rebecca Hall, Stéphanie Szostak, James Badge Dale, Jon Favreau,
and Ben Kingsley. In Iron Man 3, Tony Stark wrestles with the ramifications of the events of The Avengers during a national terrorism campaign
on the United States led by the mysterious Mandarin.
After the release of Iron Man 2 in May 2010, director Favreau chose not to return for a third film. Black was hired to write and direct
the sequel in February 2011, working with Pearce to make the script more character-centric, focus on thriller elements, and use concepts
from Warren Ellis's "Extremis" comic book story arc. The film's supporting cast, including Kingsley, Pearce, and Hall, were brought on
throughout April and May 2012. Filming took place from May 23 to December 17, 2012, primarily at EUE/Screen Gems Studios in Wilmington,
North Carolina. Additional filming took place around North Carolina as well as in Florida, Los Angeles, and China; an extended version
of the film specifically for Chinese audiences was created. Seventeen companies provided the film's visual effects.
Iron Man 3 premiered at the Grand Rex in Paris on April 14, 2013, and released in the United States on May 3, as the first film in
Phase Two of the MCU. It received praise from critics for its performances, visual effects, action sequences, humor, and Brian Tyler's
musical score, while critics and audiences gave a mixed reception to its Mandarin plot twist. The film was a box office success,
grossing over $1.2 billion worldwide making it the second highest-grossing film of 2013 and the sixteenth film to gross over $1 billion.
At the time it also became the fifth highest-grossing film of all time while its opening weekend became the sixth-highest of all time.
The film received Best Visual Effects nominations at the Academy Awards and the BAFTA Awards.
Plot
At a New Year's Eve party in 1999, Tony Stark meets scientist Maya Hansen, the inventor of an experimental regenerative treatment named
Extremis that allows recovery from crippling injuries. Disabled scientist Aldrich Killian offers them a place in his company
Advanced Idea Mechanics, but Stark rejects him. In December 2012, seven months after the battle of New York, Stark is suffering
from post-traumatic stress disorder and is having frequent panic and anxiety attacks due to his experiences during the alien invasion
and subsequent battle. Restless, he has built dozens of new Iron Man suits to cope with his insomnia, creating friction with
his girlfriend Pepper Potts.
A string of bombings claimed by a terrorist known as the Mandarin has left intelligence agencies bewildered by a lack of forensic evidence.
Stark's security chief Happy Hogan is badly injured in one such attack at the TCL Chinese Theatre, prompting Stark to boldly issue
a televised threat to the Mandarin, revealing his home address in the process. The Mandarin sends gunship helicopters to destroy Stark's home.
Hansen, who came to warn Stark, survives the attack with Potts. Stark escapes in an experimental new Iron Man suit, which his
artificial intelligence J.A.R.V.I.S. pilots to rural Tennessee, following a flight plan from Stark's investigation into the Mandarin.
Stark's new armor is not fully functional and lacks sufficient power to return to California, leaving the world to believe him dead.
Stark investigates the remains of a local explosion bearing the hallmarks of a Mandarin attack. He discovers the "bombings" were triggered
by soldiers subjected to Extremis whose bodies explosively rejected the treatment. These explosions were falsely attributed
to a terrorist plot in order to cover up Extremis's flaws. Stark witnesses Extremis first hand when Mandarin agents Savin and Brandt
attack him: Stark kills Brandt and incapacitates Savin. Meanwhile, Killian resurfaces and kidnaps Potts with assistance from Hansen.
American intelligence agencies continue to search for the Mandarin's location, with James Rhodes—the former War Machine,
now re-branded as the Iron Patriot—lured into a trap to steal his Iron-Man-like armor.
Stark traces the Mandarin to Miami and infiltrates his headquarters using improvised weapons. Inside, he discovers the Mandarin is actually
an English actor named Trevor Slattery, who is oblivious to the actions carried out in his image. Killian, who appropriated
Hansen's Extremis research as a cure for his own disability and expanded the program to include injured war veterans, reveals he is
the real Mandarin behind Slattery's cover. After capturing Stark, Killian reveals he has subjected Potts to Extremis in the hope Stark
will help fix Extremis's flaws while trying to save her. Killian kills Hansen when she tries to stop him.
Stark escapes and reunites with Rhodes, discovering that Killian intends to attack President Ellis aboard Air Force One,
using the Iron Patriot armor, controlled by Savin. Stark kills Savin, saving the passengers and crew, but cannot stop Killian
from abducting Ellis and destroying Air Force One. They trace Killian to an impounded damaged oil tanker where Killian intends
to kill Ellis on live television. The Vice President will become a puppet leader, following Killian's orders, in exchange for Extremis to cure
his young daughter's disability. On the platform, Stark works to save Potts, as Rhodes goes after the President. Stark summons his remaining Iron Man suits,
controlled remotely by J.A.R.V.I.S., to provide air support. Rhodes secures the President and takes him to safety, while Stark discovers Potts
has survived the Extremis procedure; before he can save her, a rig collapses around them and she falls to the platform below,
causing Stark to believe her dead. Stark fights Killian, but finds himself cornered. Potts, whose Extremis powers allowed her to survive her fall,
intervenes and kills Killian to save Stark.
Stark orders J.A.R.V.I.S. to remotely destroy all the Iron Man suits as a sign of his devotion to Potts, while the Vice President and Slattery
are arrested. With Stark's help, Potts's Extremis effects are stabilized; and Stark promises to scale back his life as Iron Man,
undergoing surgery to remove the shrapnel near his heart and throwing his obsolete chest arc reactor into the sea. He muses that, even without
the technology, he will always be Iron Man.
Cast
Robert Downey Jr. as Tony Stark / Iron Man:
A self-described genius, billionaire, playboy, and philanthropist with mechanical suits of armor of his own invention.
Stark now struggles to come to terms with his near-death experience in The Avengers, suffering from anxiety attacks. On making a third Iron Man film,
Downey said, "My sense of it is that we need to leave it all on the field—whatever that means in the end.
You can pick several different points of departure for that." On following up The Avengers, Downey said they "tried to be practical,
in a post-Avengers world. What are his challenges now? What are some limitations that might be placed on him? And what sort of threat would have him,
as usual, ignore those limitations?" Screenwriter Drew Pearce compared Tony to an American James Bond for both being "heroes with a sense of danger to them,
and unpredictab[le]" even if Stark was a "free agent" instead of an authority figure like Bond. He also likened Tony to the protagonists of 1970s films
such as The French Connection, where "the idiosyncrasies of the heroes is what made them exciting."
Gwyneth Paltrow as Virginia "Pepper" Potts:
Stark's girlfriend, longtime associate, and the current CEO of Stark Industries. Paltrow says of her character's relationship to Tony,
"She still adores Tony, but she absolutely gets fed up with him. He gets caught in a feedback loop." Kevin Feige commented on Pepper's role in the film:
"The love triangle in this movie is really between Tony, Pepper and the suits. Tony, Pepper and his obsession with those suits, and the obsession
with technology." Feige also stated the film uses the character to play with the damsel in distress trope, and posits the question,
"Is Pepper in danger or is Pepper the savior?"
Don Cheadle as James "Rhodey" Rhodes / Iron Patriot:
Stark's best friend and the liaison between Stark Industries and the U.S. Air Force in the department of acquisitions. Rhodes operates
the redesigned/upgraded War Machine armor, taking on an American flag-inspired color scheme similar to the Iron Patriot armor from the comics.
Feige said of Rhodes and the armor, "The notion in the movie is that a red, white and blue suit is a bold statement, and it's meant to be.
With Rhodey, he's very much the foil to Tony's eccentricities, and in this one you get to see this and be reminded of the trust and friendship
between them in that great Shane Black buddy-cop fashion." In the film, the president asks Rhodey to take up the moniker "Iron Patriot," and don the red,
white, and blue suit, in order to be the government's "American hero" in response to the events in The Avengers.
Guy Pearce as Aldrich Killian:
The creator of the Extremis virus and the founder and owner of the science and development organization Advanced Idea Mechanics who adopts the mantle
of the Mandarin as his own. Killian develops Extremis to cure his own debilitating disability; in addition to his regenerative healing qualities,
he has superhuman strength and the ability to generate extreme heat. Prolonged exposure to Extremis also grants him the ability to breathe fire.
On taking the role, Pearce said, "I feel a little more experimental in what I'll take on these days, but I still don't know that I would want to play
the superhero myself, since I'm playing a different kind of character in this film ... The main difference was that, when I did The Time Machine,
I was pretty much in all of it, so it was a really grueling experience. Prometheus and Iron Man are really kind of cameo stuff, so the experience
of shooting them ... I mean, on some level, it's tricky because you feel like a bit of an outsider. You don't really live the experience that
you do when you're there all day every day with everybody. But at the same time, it can be more fun sometimes because you're just working
in concentrated spurts." Pearce described his character as a man "who came into this world with a number of physical disabilities.
He's never been able to accept those limitations though and has spent most of his life trying to overcome them in any way he can.
His tenacity and blind determination in fighting for a better life are seen by some as irritating, as he often comes across as obnoxious.
He just won't accept the cards he was dealt, and being as intelligent as he is, has real drive to change and become a different person."
Shane Black specified, "Ultimately we do give you the Mandarin, the real guy, but it's Guy Pearce in the end with the big dragon tattooed on his chest."
He elaborated, "Do they hand me a blank check and say, 'Go break something!' Or, 'Go violate some long-standing comic book treaty that fans
have supported for years?' No, but they'll say: 'Let's break something together.' So it's okay to come up with these crazy things, these far out ideas
and they'll fly. It's just that the Marvel guys have to be in the room."
Rebecca Hall as Maya Hansen:
A geneticist whose work helped Killian create Extremis. Hall said Hansen would be a "strong female character," and described her decision to take the role,
saying, "I decided to do Iron Man 3 because I've never done the 'hurry up and wait' movie before. Even the studio movies I've done have been
small studio movies, or indie films that we made on a wing and a prayer. I love those, but Iron Man is refreshing in a way because it's something
out of my realm of experiences." Hall confirmed her character's role was greatly reduced in the final film, saying, "I signed on to do something
that was a substantial role. She wasn't entirely the villain—there have been several phases of this—but I signed on to do something very different
to what I ended up doing."
Ty Simpkins as Harley Keener:
A child who lives in Rose Hill, Tennessee. He assists Tony Stark when the latter breaks into his garage in order to repair his suit. He later helps
Stark investigate various incidents associated with the Extremis project, and afterwards he is rewarded by Stark with a garage-full of modern
engineering tools and equipment. Simpkins reprises his role as Harley in a cameo in Avengers: Endgame. Simpkins has stated he has a three-picture deal
with Marvel Studios.
Stéphanie Szostak as Ellen Brandt:
A war veteran who becomes an assassin after her exposure to Extremis. Describing Brandt, Szostak says, ". Extremis was a second chance at life.
We talked about what you feel like and I think it almost makes you a fuller version of who you are, all your weakness and your qualities—just
everything gets enhanced. I saw it as very freeing, almost you become your true-self and your fantasy-self all at once."
The writers originally envisioned Brandt as Killian's main henchman, which would return throughout the movie to fight Tony, but eventually that role
was reassigned to Eric Savin.
James Badge Dale as Eric Savin:
Killian's Extremis-powered henchman. Dale stated his character in the film was "loosely based on" the comic version of the character.
According to Dale, "Ben Kingsley is the mouthpiece. Guy Pearce is the brain. I'm the muscle."
Jon Favreau as Happy Hogan:
Tony Stark's former bodyguard and chauffeur who now serves as Stark Industries' head of security department. Favreau, who served as both actor
and director on the previous two Iron Man films, said participating in the new film was "like being a proud grandfather who doesn't have to change
the diapers but gets to play with the baby."
Ben Kingsley as Trevor Slattery:
A British actor whom Killian hired to portray the Mandarin, a terrorist persona in jammed television broadcasts in which he is depicted as the leader
of the international terrorist organization the Ten Rings.[34] Kingsley was filming Ender's Game when he was cast, and said that,
"Quite soon I'll be with everybody and we'll be discussing the look and the feel and the direction of the character. It's very early days yet,
but I'm so thrilled to be on board."On his performance, Kingsley stated: "I wanted a voice that would disconcert a Western audience.
I wanted a voice that would sound far more homegrown and familiar—a familiarity like a teacher's voice or a preacher's voice.
The rhythms and tones of an earnest, almost benign, teacher—trying to educate people for their own good."
The Mandarin was initially set to appear in the first Iron Man film, but he was put off for a sequel as the filmmakers felt that he was
"too ambitious for a first film." On the character, Feige stated, "The Mandarin is Iron Man's most famous foe in the comics mainly because he's been
around the longest. If you look, there's not necessarily a definitive Mandarin storyline in the comics. So it was really about having an idea."
Shane Black explains Ben Kingsley's Mandarin is not Chinese in the film as he is in the comics in order to avoid the Fu Manchu stereotype:
"We're not saying he's Chinese, we're saying he, in fact, draws a cloak around him of Chinese symbols and dragons because it represents
his obsessions with Sun Tzu in various ancient arts of warfare that he studied." The filmmakers also cited Colonel Kurtz from Apocalypse
Now as an influence for the character. The videos where the Mandarin gives historical background to the attacks expressed how it emerged
as the product of "a think tank of people trying to create a modern terrorist." Thus the Mandarin "represents every terrorist in a way,"
from South American insurgency tactics to the videos of Osama bin Laden.
Paul Bettany reprises his role from previous films as J.A.R.V.I.S., Stark's AI system. Ashley Hamilton portrays Taggart, one of the Extremis soldiers.
William Sadler plays President Ellis, (named after Warren Ellis, who wrote the "Extremis" comics arc that primarily influenced the film's story)
and Miguel Ferrer plays Vice President Rodriguez. Adam Pally plays Gary, a cameraman who helps Stark. Shaun Toub reprises his role
as Yinsen from the first Iron Man film in a brief cameo, and Stan Lee makes a cameo appearance as a beauty pageant judge.
Dale Dickey plays Mrs. Davis, mother of an Extremis subject that is framed as a terrorist.
Wang Xueqi briefly plays Dr. Wu in the general release version of the film.
A cut of the film produced for release exclusively in China includes additional scenes featuring Wang and an appearance by Fan Bingbing
as one of his assistants. Mark Ruffalo makes an uncredited cameo appearance, reprising his role as Bruce Banner from The Avengers,
in a post-credits scene. Comedians Bill Maher and Joan Rivers, and Fashion Police co-host George Kotsiopoulos have cameo appearances
as themselves on their respective real-world television programs, as do newscasters Josh Elliott, Megan Henderson, Pat Kiernan, and Thomas Roberts.

53
examples/nlp/conditional_generation/data/minecraft/wiki1.txt Normal file
View file

@ -0,0 +1,53 @@
Minecraft is a sandbox video game developed by Mojang Studios. The game was created by Markus Notch Persson in the Java programming language.
Following several early test versions, it was released as a paid public alpha for personal computers in 2009 before officially releasing in November 2011,
with Jens Bergensten taking over development. Minecraft has since been ported to several other platforms and is the best-selling video game of all time,
with 200 million copies sold and 126 million monthly active users as of 2020.
In Minecraft, players explore a blocky, procedurally-generated 3D world with infinite terrain, and may discover and extract raw materials, craft tools and items,
and build structures or earthworks. Depending on game mode, players can fight computer-controlled mobs, as well as cooperate with or compete against
other players in the same world. Game modes include a survival mode, in which players must acquire resources to build the world and maintain health, and a creative mode,
where players have unlimited resources. Players can modify the game to create new gameplay mechanics, items, and assets.
Minecraft has been critically acclaimed, winning several awards and being cited as one of the greatest video games of all time.
Social media, parodies, adaptations, merchandise, and the annual MineCon conventions played large roles in popularizing the game.
It has also been used in educational environments, especially in the realm of computing systems, as virtual computers and hardware devices have been built in it.
In 2014, Mojang and the Minecraft intellectual property were purchased by Microsoft for US$2.5 billion.
A number of spin-off games have also been produced, such as Minecraft: Story Mode, Minecraft Dungeons, and Minecraft Earth.
Minecraft is a 3D sandbox game that has no specific goals to accomplish, allowing players a large amount of freedom in choosing how to play the game.
However, there is an achievement system, known as advancements in the Java Edition of the game.
The game world is composed of rough 3D objects, mainly cubes and fluids, and commonly called blocks representing various materials,
such as dirt, stone, ores, tree trunks, water, and lava. The core gameplay revolves around picking up and placing these objects.
These blocks are arranged in a 3D grid, while players can move freely around the world. Players can mine blocks and then place them elsewhere,
enabling them to build things.
Many commentators have described the game's physics system as unrealistic;
liquids continuously flow for a limited horizontal distance from source blocks, which can be removed by placing a solid block in its place or by scooping it into a bucket.
The game also contains a material known as redstone, which can be used to make primitive mechanical devices, electrical circuits, and logic gates,
allowing for the construction of many complex systems.
The game world is virtually infinite and procedurally generated as players explore it, using a map seed that is obtained from the system clock at the time of world creation
or manually specified by the player. There are limits on vertical movement, but Minecraft allows an infinitely large game world to be generated on the horizontal plane.
Due to technical problems when extremely distant locations are reached, however, there is a barrier preventing players from traversing to locations beyond
30,000,000 blocks from the center. The game achieves this by splitting the world data into smaller sections called chunks that are only created or loaded
when players are nearby. The world is divided into biomes ranging from deserts to jungles to snowfields; the terrain includes plains, mountains, forests, caves,
and various lava water bodies. The in-game time system follows a day and night cycle, and one full cycle lasts 20 realtime minutes.
When starting a new world, players must choose one of five game modes, as well as one of four difficulties, ranging from peaceful to hard.
Increasing the difficulty of the game causes the player to take more damage from mobs, as well as having other difficulty-specific effects.
For example, the peaceful difficulty prevents hostile mobs from spawning, and the hard difficulty allows players to starve to death if their hunger bar is depleted.
Once selected, the difficulty can be changed, but the game mode is locked and can only be changed with cheats.
New players have a randomly selected default character skin of either Steve or Alex, but the option to create custom skins was made available in 2010.
Players encounter various non-player characters known as mobs, such as animals, villagers, and hostile creatures.
Passive mobs, such as cows, pigs, and chickens, can be hunted for food and crafting materials. They spawn in the daytime, while hostile mobs including large spiders,
skeletons, and zombies spawn during nighttime or in dark places such as caves. Some hostile mobs, such as zombies, skeletons and drowned (underwater versions of zombies),
burn under the sun if they have no headgear. Other creatures unique to Minecraft include the creeper (an exploding creature that sneaks up on the player)
and the enderman (a creature with the ability to teleport as well as pick up and place blocks).
There are also variants of mobs that spawn in different conditions; for example, zombies have husk variants that spawn in deserts.
Minecraft has two alternative dimensions besides the overworld (the main world): the Nether and the End. The Nether is a hell-like dimension accessed
via player-built portals; it contains many unique resources and can be used to travel great distances in the overworld, due to every block traveled in the Nether
being equivalent to 8 blocks traveled in the overworld. The player can build an optional boss mob called the Wither out of materials found in the Nether.
The End is a barren land consisting of many islands. A boss dragon called the Ender Dragon dwells on the main island.
Killing the dragon opens access to an exit portal, which upon entering cues the game's ending credits and a poem written by Irish novelist Julian Gough.
Players are then teleported back to their spawn point and may continue the game indefinitely.

95
examples/nlp/conditional_generation/data/minecraft/wiki2.txt Normal file
View file

@ -0,0 +1,95 @@
Minecraft is an indie survival sandbox construction video game designed by Markus Notch Persson and Mojang Studios and developed on Java.
The game allows players to collect, assemble, or destroy a variety of blocks in a three-dimensional, procedurally-generated environment.
The game takes inspiration from multiple sources, specifically the game Infiniminer.
Development began in May 2009, with pre-orders for the full game starting June 13, 2009. The game was officially released by Notch in mid-November 2011 during MineCon.
During late 2019, Minecraft has sold over 176 million copies worldwide across all platforms, making it the 2nd highest selling video game of all time, only behind Tetris.
Blocks make up the Minecraft landscape and are integral to gameplay. They can be broken, collected, and placed in an infinite number of arrangements.
They can be used in crafting and manipulated in many other ways. The face of a block is 16 by 16 pixels, and each block is proportionately one cubic meter.
Minecraft is a game that has no specific goals to accomplish, allowing players a large amount of freedom in choosing how to play it.
However, there is an achievements system. While the gameplay revolves through the players' experiences, the game encourages many players to choose their own way of adventure.
There is no right or wrong way to play the game, players can play the game however they want it to be.
There are two game modes in the game: Survival and Creative. In Survival mode, players must gather resources from the Overworld, hunt for food to replenish health and hunger,
and protect themselves from the deadly hostile mobs that attack them during nighttime, or in cave systems. T
he most prominent objective in Survival mode, is to mine diamonds to upgrade certain weapons or tools or make convenient armors that would help them throughout their journeys.
In Creative mode, players are free to roam in the world wherever they want, and just simply build anything that comes within their imagination, like cities,
redstone machines, etc.
The gameplay is in the first-person perspective by default, but players have the option for a third-person perspective.
The game also contains a material known as redstone, which can be used to make mechanical devices, electrical circuits, and logic gates, allowing for the construction
of many complex systems. Players can select any customized skin whenever they want to, but new players start with the default character skin of either Steve or Alex.
Players encounter various non-player characters known as mobs, such as animals, villagers, and hostile creatures. Passive mobs, such as cows, pigs, and chickens,
can be hunted for food and crafting materials. They spawn in the daytime, while hostile mobs, including large spiders, skeletons, and zombies—spawn during nighttime
or in dark places such as caves. Some hostile mobs, such as the Zombie, Skeleton, and Drowned (underwater versions of Zombies), burn under the sun if they have no headgear.
Other creatures unique to Minecraft include the Creeper (an exploding creature that sneaks up on the player) and the Enderman (a creature with the ability to teleport,
pick up, and place blocks). There are also variants of mobs that spawn in different conditions; for example, Zombies have Husk variants that spawn in deserts.
Minecraft has two alternate dimensions besides the overworld (the main world): the Nether and the End. The Nether is a hell-like dimension accessed via player-built portals;
it contains many unique resources and can be used to travel great distances in the overworld. The player can build an optional boss mob called the Wither out of materials
found in the Nether. The End is a barren land consisting of many islands. A boss dragon called the Ender Dragon dwells on the main island.
Killing the dragon opens access to an exit portal, which upon entering cues the game's ending credits and a poem written by Irish novelist Julian Gough.
There is also another portal that access the outer islands of The End, in which players could access the outer islands. Players are then allowed to teleport back
to their original spawn point in the Overworld and continue the game indefinitely.
Survival game mode.
The original and core gameplay element of the game. In this mode, players have to gather natural resources such as wood and stone found in the environment in order
to craft certain blocks and items. Depending on the difficulty, monsters spawn in darker areas outside a certain radius of the character, requiring players to build
a shelter at night. Players are required to fight off the hostile mobs that attack them in order to protect themselves from certain damages.
The mode also has a health bar which is depleted by attacks from monsters, falls, drowning, falling into lava, suffocation, starvation, and other events.
Players also have a hunger bar, which must be periodically refilled by eating food in-game, except in peaceful difficulty. If the hunger bar is depleted,
automatic healing will stop and eventually health will deplete. Health replenishes when players have a nearly full hunger bar or continuously on peaceful difficulty.
Players may also trade goods with Villagers or Piglins NPCs throughout a trading or bartering system, which involves trading emeralds or gold for different goods and vice versa.
Players may acquire experience points by killing mobs and other players, mining, smelting ores, breeding animals, and cooking food.
Experience can then be spent on enchanting tools, armor and weapons. Enchanted items are generally more powerful, last longer, or have other special effects.
Crafting.
Since the uses of natural blocks are limited, it is imperative that players make use of the crafting interface to survive. Players can combine items in a 2x2 interface
(extended to 3x3 by use of a crafting table), with each item taking up one square of the grid. Different arrangements of items create usable items,
which players can exploit to further their collection of resources, and in turn create more items.
Players can craft a wide variety of items in Minecraft. Craftable items include armor, which mitigates damage from attacks; weapons (such as swords),
which allows monsters and animals to be killed more easily; and tools, which break certain types of blocks more quickly. Some items have multiple tiers depending
on the material used to craft them, with higher-tier items being more effective and durable. Players can construct furnaces, which can cook food, process ores,
and convert materials into other materials.
Inventory.
The game has an inventory system, allowing players to carry a limited number of items. Upon dying, items in the players' inventories are dropped,
and players re-spawn at their spawn point, which by default is where players first spawn in the game, and can be reset by sleeping in a bed or using a respawn anchor.
Dropped items can be recovered if players can reach them before they disappear, or despawn, after 5 minutes.
Objectives.
While they are no specific objectives in the game, The main goals of Survival, is to simply survive, gather important resources, construct and expand your
shelter before nighttime, explore your world, have fun, upgrade your tools, materials or equipments, find diamonds, protect their companions, and fight monsters
that try to attack them. Survival also has an ending, that is optional to players, and that is to kill the hostile boss mob, Ender Dragon.
Players can also build the Wither boss by obtaining the right materials from the nether, and defeat it in order to craft a beacon.
They also have the option to protect the villagers from an pillager raid, which is triggered by the player itself.
Progress can be made once the player does all of these. Players can also beat the game's achievements, or advancements, but this is completely optional.
Creative game mode.
In creative mode, players have access to all resources and items in the game through the inventory menu, and can place or remove them instantly.
Players can toggle the ability to fly freely around the game world at will, and their characters do not take any damage and are not affected by hunger.
The game mode helps players focus on building and creating projects of any size without disturbance.
Hardcore game mode.
Hardcore mode is a survival mode variant that is locked to the hardest setting and has permadeath. If a player dies in a hardcore world,
they are no longer allowed to interact with it, so they can either be put into spectator mode and explore the world or return to title screen.
You cant change difficulty if you set it to hardcore once unlike survival mode.
Music.
Minecraft's music and sound effects were produced by German musician Daniel Rosenfeld (born: May 9, 1989 [age 31]), better known as C418.
The background music in Minecraft is simply instrumental ambient music. On 4 March 2011, Rosenfeld released a soundtrack, titled Minecraft – Volume Alpha;
it includes most of the tracks featured in Minecraft, as well as other music not featured in the game.
Kirk Hamilton of Kotaku chose the music in Minecraft as one of the best video game soundtracks of 2011. On 9 November 2013, Rosenfeld released the second official soundtrack,
titled Minecraft – Volume Beta, which includes the music that was added in later versions of the game.
A physical release of Volume Alpha, consisting of CDs, black vinyl, and limited-edition transparent green vinyl LPs, was issued by indie electronic
label Ghostly International on 21 August 2015.
Multiplayer.
Multiplayer in Minecraft is available through direct game-to-game multiplayer, LAN play, local split screen, and servers (player-hosted and business-hosted).
It enables multiple players to interact and communicate with each other on a single world. Players can run their own servers, use a hosting provider,
or connect directly to another player's game via Xbox Live. Single-player worlds have local area network support, allowing players to join a world on locally
interconnected computers without a server setup. Minecraft multiplayer servers are guided by server operators, who have access to server commands such as setting
the time of day and teleporting players. Operators can also set up restrictions concerning which usernames or IP addresses are allowed or disallowed to enter the server.
Multiplayer servers have a wide range of activities, with some servers having their own unique rules and customs. The largest and most popular server is Hypixel,
which has been visited by over 14 million unique players. Player versus player combat (PvP) can be enabled to allow fighting between players.
Many servers have custom plugins that allow actions that are not normally possible.

113
examples/nlp/conditional_generation/data/minecraft/wiki3.txt Normal file
View file

@ -0,0 +1,113 @@
Minecraft is a sandbox construction video game originally created by Markus Notch Persson. It is maintained by Mojang Studios, a part of Xbox Game Studios,
which in turn is part of Microsoft.
From its creation, Minecraft was developed almost exclusively by Notch until Jens Bergensten started working with him, and has since become head of its development.
It features music by Daniel Rosenfeld and paintings by Kristoffer Zetterstrand. Initially released as what is now known as Minecraft Classic on May 17, 2009,
the game was fully released on November 18, 2011. Since its release, Minecraft has expanded to mobile devices and consoles. On November 6, 2014,
Minecraft and all of Mojang Studios' assets were acquired by Microsoft for US$2.5 billion. Notch has since left Mojang, and is no longer working on Minecraft.
Minecraft focuses on allowing the player to explore, interact with, and modify a dynamically-generated map made of one-cubic-meter-sized blocks.
In addition to blocks, the environment features plants, mobs, and items. Some activities in the game include mining for ore, fighting hostile mobs, and crafting new blocks
and tools by gathering various resources found in the game. The game's open-ended model allows players to create structures, creations, and artwork
on various multiplayer servers or their single-player maps. Other features include redstone circuits for logic computations and remote actions,
minecarts and tracks, and a mysterious underworld called the Nether. A designated but completely optional goal of the game is to travel to a dimension called the End,
and defeat the ender dragon.
Player.
The player is the person that the user controls in the world. When the user starts a game, the player is put in a world, generated by a random or specified seed,
with an empty inventory. If the bonus chest option is enabled, a chest filled with basic items generates near the player.
The player has a health bar with 10 hearts, and can be damaged by falls, suffocation, drowning, fire, lava, lightning, cacti, sweet berry bushes,
falling into the Void, falling anvils and being hit by mobs and other players. Damage to health can be mitigated by armor or Resistance potion and health
can be restored by eating food and drinking specific potions, or if difficulty is set to Peaceful, health regenerates on its own.
Hunger is also a factor if the difficulty is not set to Peaceful, depleting over time and even faster while sprinting, jumping or swimming.
Food replenishes the hunger level, however, eating rotten flesh and raw chicken has a chance of giving the player a hunger effect.
Depending on the difficulty level, a low hunger level depletes a player's health.
A player can change their skin on the profile page of Minecraft.net or in the launcher.
Blocks.
The world of Minecraft takes place within a three-dimensional grid of cubes, with each cube being occupied by a certain type of block (not all of which are necessarily cubic).
here are different types of blocks; natural blocks such as grass, stone, and ores are randomly generated within the world. There are also blocks that players can craft,
such as a crafting table and a furnace. Resources can be extracted from blocks by hand or by using tools. Some of these resources are simply blocks in the player's
inventory that can be placed elsewhere, while others are used as material to create other blocks or tools. Others yield no practical use whatsoever.
Some blocks cannot be broken through normal survival means, e.g. bedrock, end portal frames, command blocks, and barriers.
Mining.
Various ores (in proximity of lava) that can be mined with a pickaxe.
Mining is one of the main aspects of Minecraft and is done to extract ore and other materials mainly from below the surface of the map.
These ores include coal, iron, gold, redstone, diamond, lapis lazuli, and emerald. Mining can involve digging a hole from the surface or going down through a cave.
Mineshafts create extra areas that may contain resources, since they are usually rich in ores.
Crafting and Smelting.
A crafting table, used to create most of the blocks and items in Minecraft.
Crafting allows players to create new tools and blocks using items from their inventory. Subsequent versions often contain crafting recipes for new blocks and items.
To craft, a player can use the 2×2 grid in the inventory or the 3×3 grid provided by a crafting table. Smelting requires a furnace in addition to fuel,
and processes blocks into a more useful form such as iron ores into iron ingots.
Brewing and Enchanting.
An Enchantment Table with glyphs being absorbed into it.
Brewing creates potions from various ingredients and water using a brewing stand. They are stored in a glass bottle and then consumed by the player or thrown
at other mobs to generate a certain effect based on the ingredients used to create the potion. Enchanting is also used to upgrade armor,
tools, or weapons with an enchanting table. More powerful enchantments can be accessed by gaining experience and placing bookshelves around the enchanting table.
Mobs.
A creeper in a forest. Creepers stalk the player and then explode once they get near.
Mobs (short for mobiles) are the animals and other creatures that inhabit the map. Hostile mobs attack the player while passive mobs do not. Neutral mobs attack when provoked.
The Overworld contains many passive mobs that may be killed for food or bred with one another; these include:
Pigs: drop porkchops upon death and can be ridden using a saddle.
Cows: drop beef upon death and can be milked using a bucket.
Sheep: drop mutton and 1 wool upon death and can be shorn to produce 1–3 wool.
Chickens: drop chicken meat and feathers upon death and lay eggs.
Horses: drop leather upon death and can be ridden using a saddle, traveling much faster than pigs.
Bats: ambient mobs that fly around caves.
Common hostile mobs found throughout the Overworld include:
Zombies: attack by melee damage.
Skeletons: have a bow and infinitely many arrows.
Spiders: jump large distances and can climb walls.
Witches: use potions.
Creepers: explode when near the player.
Endermen: tall, black creatures with purple eyes and turn aggressive when the player looks at them.
The Overworld also contains some rarer mobs that spawn only on occasion or in specific biomes:
Spider jockeys: a skeleton riding a spider.
Chicken jockeys: a baby zombie riding a chicken.
Slimes: spawn deep within the map and in swamplands.
Villagers: inhabit villages and can trade with the player.
Parrots: can imitate the sounds of nearby mobs.
Wolves: can be tamed by the player and attack enemy mobs if the player engages or is attacked by them
Mooshrooms: mushroom variants of cows that spawn in mushroom field biomes.
Vindicator: spawn in woodland mansions
Pillager: spawn in pillager outposts
Evoker: spawn in Woodland Mansions
Some mobs can be found exclusively in the Nether, including:
Ghasts: flying mobs that shoot exploding fireballs at the player.
Zombified piglins: wield golden swords and attack in hordes if provoked.
Wither skeletons: tall, black variants of regular skeletons that wield stone swords and drop coal and, occasionally, wither skeleton skulls that
can be used to summon the wither.
Blazes: shoot fireballs at players and hover above the ground.
Magma cubes: similar to Overworld slimes.
The End contains the ender dragon, which is the main boss mob in Minecraft and allows the player to exit back to the Overworld when it dies.
Withers are the second boss mob in Minecraft, and are created by the player by placing wither skeleton skulls on top of soul sand in a specific pattern.
When spawned, they shoot wither skulls at nearby non-undead mobs.
The Nether.
The Nether is a dimension in Minecraft, accessible from the Overworld by a nether portal. It consists of five biomes, those being the Nether Wastes,
the Basalt Deltas, the Crimson and Warped Forests and the Soul Sand Valleys: each biome has unique generation and terrain.
It is populated by zombified piglins, blazes, ghasts, wither skeletons, magma cubes, piglins and hoglins.
The End.
The End is another dimension of the game where the player battles the ender dragon. The End is accessible by entering an end portal found in a stronghold.
The End is composed of end stone and is inhabited by endermen. It also contains tall obsidian pillars on top of which are end crystals that heal the ender dragon.
Once the ender dragon is slain, the exit portal is created in the center of the map, and an end gateway portal is created near an edge of the map,
which transports the player to the expansive outer end islands.
Multiplayer.
Minecraft multiplayer servers have developed to include their own rules and customs, guided by their administrators and moderators.
The term griefer, meaning a player who causes grief, is a typical term on the Internet but has taken up its definition on Minecraft servers: a person who destroys
or defiles other users' creations on servers. Griefers are the reason many server administrators make rules, but this has been taken a step further with modifications
to the Minecraft server and even plugin-based replacement servers such as Bukkit. Because of these plugin-based servers, new user-created features have shown
up in Minecraft. This includes features like money, vehicles, protection, RPG elements and more. These features normally do not require modification to a user's
client and can be accessed by using chat commands. With the default controls, the chat screen is brought up by pressing T.
One popular game on multiplayer servers is Spleef (a play on the word grief), a game where the player aims to make another player drop through the floor
by destroying blocks beneath the opponent's feet. This is typically played in a designated area and is usually run automatically using server plugins.
Many popular multiplayer servers exist that may contain custom minigames or large Survival or Creative worlds.

65
examples/nlp/conditional_generation/lfqa_utils.py
View file

@ -50,41 +50,50 @@ def load_data(directory, max_size):
for chunk in text_chunks:
print(f'{len(chunk.split())} - {chunk}')
print('-----------------------------')
return text_chunks
qa_url = "http://34.72.18.118/qa_api"
def extractive_qa(question, content):
payload = f'{{ \
"question": "{question}", \
"input_context": "{content}" \
}}'
def extractive_qa(question, context):
'''
Uses Extractive QA server.
Input: question and context
Response: short span with the answer and full sentence that contains this span or 'Not found'
'''
url = "http://35.193.226.52:9000/longformqa"
payload = (
f'{{\r\n"question": "{question}",\r\n\"context\": "{context}",\r\n'
f'"maxLength": 128,\r\n"minLength": 1,\r\n "numBeams": 8\r\n}}'
)
headers = {
'Content-Type': "application/json",
'User-Agent': "PostmanRuntime/7.18.0",
'Accept': "*/*",
'Cache-Control': "no-cache",
'Postman-Token': "30549445-ea95-416a-bd74-8ab14ad2036d,0dd97345-55e6-4517-970c-f7032bd1a0d3",
'Host': "35.193.226.52:9000",
'Accept-Encoding': "gzip, deflate",
'Content-Length': "1513",
'Connection': "keep-alive",
'cache-control': "no-cache",
'Content-Type': 'application/json',
}
print(payload)
response = requests.request("POST", url, data=payload, headers=headers)
if response.ok == True:
print(response.text)
response_dict = json.loads(response.text)
return (response_dict['result_eqa'], response_dict['context'])
response = requests.request("POST", qa_url, headers=headers, data=payload)
# print(response.text.encode('utf8'))
response_dict = json.loads(response.text)
if response_dict['result']:
#print(response_dict['result'])
return (response_dict['result'], response_dict['context'])
else:
#print("Answer not found in the context")
return ('Not found', 'Not found')
qa_url = "http://34.72.18.118/gen_qa"
def answer_extender(question, content):
payload = f'{{ \
"question": "{question}", \
"input_context": "{content}" \
}}'
headers = {
'Content-Type': 'application/json',
}
response = requests.request("POST", qa_url, headers=headers, data=payload)
response_dict = json.loads(response.text)
if response_dict['result']:
return (response_dict['result'], response_dict['context'])
else:
print("Answer not found in the context")
return ('Not found', 'Not found')

61
examples/nlp/conditional_generation/qa_demo.py
View file

@ -38,33 +38,32 @@ def load_models():
@st.cache(allow_output_mutation=True)
def load_indexes(qar_model, qar_tokenizer):
marvel_snippets = load_data("/home/vgetselevich/data/marvel/wiki/", 100)
snippets = load_data("/home/vgetselevich/data/minecraft/", 100)
# prepare IR index
if not os.path.isfile('marvel_passages_reps_32_l-8_h-768_b-512-512.dat'):
if not os.path.isfile('minecraft_passages_reps_32_l-8_h-768_b-512-512.dat'):
print("*** Generating dense index ***")
make_qa_dense_index_text_chunks(
qar_model,
qar_tokenizer,
# wiki40b_snippets,
marvel_snippets,
snippets,
device='cuda:0',
index_name='marvel_passages_reps_32_l-8_h-768_b-512-512.dat',
index_name='minecraft_passages_reps_32_l-8_h-768_b-512-512.dat',
)
faiss_res = faiss.StandardGpuResources()
# wiki40b_passages = nlp.load_dataset(path="wiki_snippets", name="wiki40b_en_100_0")["train"]
marvel_passage_reps = np.memmap(
"marvel_passages_reps_32_l-8_h-768_b-512-512.dat",
"minecraft_passages_reps_32_l-8_h-768_b-512-512.dat",
dtype="float32",
mode="r",
shape=(len(marvel_snippets), 128),
shape=(len(snippets), 128),
)
wiki40b_index_flat = faiss.IndexFlatIP(128)
wiki40b_gpu_index_flat = faiss.index_cpu_to_gpu(faiss_res, 1, wiki40b_index_flat)
wiki40b_gpu_index_flat.add(marvel_passage_reps) # TODO fix for larger GPU
return (marvel_snippets, wiki40b_gpu_index_flat)
return (snippets, wiki40b_gpu_index_flat)
qar_tokenizer, qar_model, s2s_tokenizer, s2s_model = load_models()
@ -111,21 +110,21 @@ def answer_question(
return (answer, support_list)
st.title("Marvel Universe Question Answering")
st.title("Minecraft Question Answering")
# Start sidebar
# header_html = "<img src='https://huggingface.co/front/assets/huggingface_logo.svg'>"
# header_html = "<img src='iron_man.png'>"
header_html = "<img width='150' height='200' src='https://vignette.wikia.nocookie.net/marvelcinematicuniverse/images/f/f2/Iron_Man_Armor_-_Mark_LXXXV.png/revision/latest/top-crop/width/360/height/450?cb=20190401222437'>"
header_html = "<img width='200' height='150' src='https://cdn.vox-cdn.com/thumbor/auxLh0jMJyEywGb5SyA02trm6ag=/1400x1400/filters:format(jpeg)/cdn.vox-cdn.com/uploads/chorus_asset/file/4093800/image.0.jpg'>"
header_full = """
<html>
<head>
<style>
.img-container {
padding-left: 70px;
padding-right: 70px;
padding-top: 50px;
padding-bottom: 50px;
padding-left: 50px;
padding-right: 100px;
padding-top: 100px;
padding-bottom: 100px;
background-color: #f0f3f9;
}
</style>
@ -143,11 +142,11 @@ st.sidebar.markdown(
header_full, unsafe_allow_html=True,
)
# Generative QA on Marvel Universe
# Generative QA on Minecraft data
description = """
---
This demo presents generative and extractive answers to Marvel Universe content questions.
First dense IR model fetches a set of relevant snippets from Marvel wiki docs given the question,
This demo presents generative and extractive answers to Minecraft content questions.
First dense IR model fetches a set of relevant snippets from Minecraft wiki docs given the question,
and then Bart model use them as a context to generate and answer and
extractive QA model creates an alternative answer.
"""
@ -222,9 +221,9 @@ if generate_options:
# start main text
questions_list = [
"<MY QUESTION>",
"Who is iron man?",
"who is Tony Stark?",
"who directed iron man?",
"what is minecraft?",
"who created minecraft game?",
"which modes minecraft game has?",
]
question_s = st.selectbox(
"What would you like to ask? ---- select <MY QUESTION> to enter a new query", questions_list, index=1,
@ -237,8 +236,8 @@ else:
if st.button("Show me!"):
if action in [0, 1, 3]:
if index_type == "mixed":
_, support_list_dense = make_support(question, source=wiki_source, method="dense", n_results=10)
_, support_list_sparse = make_support(question, source=wiki_source, method="sparse", n_results=10)
_, support_list_dense = make_support(question, source=wiki_source, method="dense", n_results=5)
_, support_list_sparse = make_support(question, source=wiki_source, method="sparse", n_results=5)
support_list = []
for res_d, res_s in zip(support_list_dense, support_list_sparse):
if tuple(res_d) not in support_list:
@ -248,7 +247,7 @@ if st.button("Show me!"):
support_list = support_list[:10]
question_doc = "<P> " + " <P> ".join([res[-1] for res in support_list])
else:
question_doc, support_list = make_support(question, source=wiki_source, method=index_type, n_results=10)
question_doc, support_list = make_support(question, source=wiki_source, method=index_type, n_results=5)
if action in [0, 3]:
# Geenerative QA
answer, support_list = answer_question(
@ -263,17 +262,17 @@ if st.button("Show me!"):
temp=temp,
)
# Extractive QA
ex_answer_span, ex_answer_sent = extractive_qa(question, question_doc.replace("<P> ", ""))
st.markdown("### The model generated answer is:")
st.markdown("### The Generative Answer is:")
st.write(answer)
st.markdown("### The model extractive answer is:")
st.write(ex_answer_span + " - " + ex_answer_sent)
# Extractive QA
ex_answer_span, ex_answer_context = extractive_qa(question, question_doc.replace("<P> ", ""))
st.markdown("### The Extractive Answer is:")
st.write(ex_answer_span)
if action in [0, 1, 3] and wiki_source != "none":
st.markdown("--- \n ### The model is drawing information from the following Marvel snippets:")
st.markdown("--- \n ### The model is drawing information from the following Minecraft snippets:")
for i, res in enumerate(support_list):
wiki_url = "https://en.wikipedia.org/wiki/{}".format(res[0].replace(" ", "_"))
sec_titles = res[1].strip()

116
examples/nlp/conditional_generation/server.py Normal file
View file

@ -0,0 +1,116 @@
from lfqa_utils import *
import socket
HOST = '10.110.42.102'
PORT = 25556
def load_models():
qar_tokenizer = AutoTokenizer.from_pretrained("yjernite/retribert-base-uncased")
qar_model = AutoModel.from_pretrained("yjernite/retribert-base-uncased").to("cuda:0")
_ = qar_model.eval()
s2s_tokenizer = AutoTokenizer.from_pretrained("yjernite/bart_eli5")
s2s_model = AutoModelForSeq2SeqLM.from_pretrained("yjernite/bart_eli5").to("cuda:0")
# save_dict = torch.load("seq2seq_models/eli5_bart_model_blm_2.pth")
# s2s_model.load_state_dict(save_dict["model"])
_ = s2s_model.eval()
return (qar_tokenizer, qar_model, s2s_tokenizer, s2s_model)
def load_indexes(qar_model, qar_tokenizer):
snippets = load_data("/home/vgetselevich/data/minecraft/", 100)
# prepare IR index
if not os.path.isfile('minecraft_passages_reps_32_l-8_h-768_b-512-512.dat'):
print("*** Generating dense index ***")
make_qa_dense_index_text_chunks(
qar_model,
qar_tokenizer,
snippets,
device='cuda:0',
index_name='minecraft_passages_reps_32_l-8_h-768_b-512-512.dat',
)
faiss_res = faiss.StandardGpuResources()
# wiki40b_passages = nlp.load_dataset(path="wiki_snippets", name="wiki40b_en_100_0")["train"]
passage_reps = np.memmap(
"minecraft_passages_reps_32_l-8_h-768_b-512-512.dat",
dtype="float32",
mode="r",
shape=(len(snippets), 128),
)
wiki40b_index_flat = faiss.IndexFlatIP(128)
wiki40b_gpu_index_flat = faiss.index_cpu_to_gpu(faiss_res, 1, wiki40b_index_flat)
wiki40b_gpu_index_flat.add(passage_reps) # TODO fix for larger GPU
return (snippets, wiki40b_gpu_index_flat)
def get_answer(question, qar_tokenizer, qar_model, qa_s2s_tokenizer, qa_s2s_model, snippets, gpu_dense_index, use_eqa=True):
# create support document with the dense index
doc, res_list = query_qa_dense_index(
question, qar_model, qar_tokenizer, snippets, gpu_dense_index, n_results=5, device='cuda:0'
)
# concatenate question and support document into BART input
question_doc = "question: {} context: {}".format(question, doc)
# generate an answer with beam search
gen_answer = qa_s2s_generate(
question_doc,
qa_s2s_model,
qa_s2s_tokenizer,
num_answers=1,
num_beams=8,
min_len=10,
max_len=50,
max_input_length=1024,
device="cuda:0",
)[0]
print(question)
print("Generative Answer: " + gen_answer)
print("Context: " + doc.replace("<P> ", ""))
ex_answer_span = ''
if use_eqa:
ex_answer_span, ex_answer_context = extractive_qa(question, doc.replace("<P> ", ""))
print("Extractive Answer: " + ex_answer_span)
return (gen_answer, ex_answer_span)
def main():
qar_tokenizer, qar_model, qa_s2s_tokenizer, qa_s2s_model = load_models()
snippets, gpu_dense_index = load_indexes(qar_model, qar_tokenizer)
with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
s.bind((HOST, PORT))
print(f"IR + QA server is running on: {HOST}:{PORT}")
s.listen()
while True:
conn, addr = s.accept()
with conn:
print('Connected by', addr)
while True:
data = conn.recv(1024).decode()
if not data:
break
gen_answer, ex_answer_span = get_answer(data, qar_tokenizer, qar_model, qa_s2s_tokenizer,
qa_s2s_model, snippets, gpu_dense_index, use_eqa=False)
if ex_answer_span:
conn.sendall(ex_answer_span.encode('utf-8'))
else:
conn.sendall(gen_answer.encode('utf-8'))
# response = f'Extractive: {ex_answer_span} Generative: {gen_answer}'
# conn.sendall(response.encode('utf-8'))
if __name__ == '__main__':
main()