liujch1998/crystal-11b

Model Card for Crystal

Crystal is an introspective reasoning model commonsense QA. See our paper at: https://arxiv.org/abs/2310.04921.

Model Details

Model Description

Crystal can answer a given commonsense question by first generating a relevant knowledge statement, and then predict the final answer by referencing the generated knowledge. We call this process "introspective reasoning", and it improves both the prediction accuracy and the interpretability of neural models at reasoning tasks.

• Developed by: Jiacheng Liu, Ramakanth Pasunuru, Hannaneh Hajishirzi, Yejin Choi, Asli Celikyilmaz
• Shared by [optional]: Jiacheng Liu
• Model type: Transformers
• Language(s) (NLP): English
• License: MIT
• Finetuned from model [optional]: t5-11b

Model Sources [optional]

• Repository: https://github.com/liujch1998/crystal
• Paper [optional]: https://arxiv.org/abs/2310.04921
• Demo [optional]: https://huggingface.co/spaces/liujch1998/crystal

Uses

Direct Use

Crystal is intended to answer commonsense questions via an "introspective reasoning" process.

Out-of-Scope Use

Crystal is a research prototype and may give incorrect answers or reasoning process. Do not use for making critical decisions. It is intended to answer questions about commonsense, and may be unreliable when taking input out of this scope.

Bias, Risks, and Limitations

See the Limitations section of our paper.

Recommendations

Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.

How to Get Started with the Model

Use the code below to get started with the model.

from transformers import T5Tokenizer, T5ForConditionalGeneration

tokenizer = AutoTokenizer.from_pretrained('liujch1998/crystal-11b')
model = AutoModelForSeq2SeqLM.from_pretrained('liujch1998/crystal-11b')
model.eval()

max_question_len, max_knowledge_len, max_answer_len = 128, 32, 2
k = 1 # number of knowledge statements to generate
top_p = 0.0001

question = 'If the mass of an object gets bigger what will happen to the amount of matter contained within it? \\n (A) gets bigger (B) gets smaller'
choices = ['A', 'B']

choices_ids = tokenizer(choices, return_tensors='pt', padding='max_length', truncation='longest_first', max_length=max_answer_len).input_ids # (C, AL)

prompt = question + ' \\n Knowledge: '
prompt_tok = tokenizer(prompt, return_tensors='pt', padding='max_length', truncation='longest_first', max_length=max_question_len) # (1, QL)
knowledges_ids = self.model.generate(
    input_ids=prompt_tok.input_ids,
    attention_mask=prompt_tok.attention_mask,
    max_length=max_knowledge_len + 1,
    min_length=3,
    do_sample=True,
    num_return_sequences=k,
    top_p=top_p,
) # (K, KL); begins with 0 ([BOS]); ends with 1 ([EOS])
knowledges_ids = knowledges_ids[:, 1:].contiguous() # no beginning; ends with 1 ([EOS])
knowledges = tokenizer.batch_decode(knowledges_ids, skip_special_tokens=True, clean_up_tokenization_spaces=True)

prompts = [question + (f' \\n Knowledge: {knowledge} \\n Answer: ' if knowledge != '' else ' \\n Answer:') for knowledge in knowledges]
prompts_tok = self.tokenizer(prompts, return_tensors='pt', padding='max_length', truncation='longest_first', max_length=max_question_len + max_knowledge_len) # (K, QL+KL)
output = model(
    input_ids=prompts_tok.input_ids,
    attention_mask=prompts_tok.attention_mask,
    labels=choices_ids[0].unsqueeze(0).repeat(len(knowledges), 1),
)
logitsss = output.logits # (K, AL, V)
logitss = logitsss[:, 0, :] # (K, V)
choice_ids = choices_ids[:, 0] # (C)
answer_logitss = logitss.gather(dim=1, index=choice_ids.unsqueeze(0).expand(len(knowledges), -1)) # (K, C)
answer_probss = answer_logitss.softmax(dim=1) # (K, C)
answer_probs = answer_probss.max(dim=0).values # (C)
pred = answer_probs.argmax(dim=0).item()
pred = choices[pred]

print(f'Question: {question}\nKnowledge: {knowledges[0]}\nAnswer: {pred}')

You may also refer to https://huggingface.co/spaces/liujch1998/crystal/blob/main/app.py#L10-L86 for implementation.

Citation [optional]

BibTeX:

@article{Liu2023CrystalIR,
  title={Crystal: Introspective Reasoners Reinforced with Self-Feedback},
  author={Jiacheng Liu and Ramakanth Pasunuru and Hannaneh Hajishirzi and Yejin Choi and Asli Celikyilmaz},
  journal={ArXiv},
  year={2023},
  volume={abs/2310.04921}
}

Model Card Contact

Jiacheng Liu