relevancy.py

"""
run:
python -m relevancy run_all_day_paper \
  --output_dir ./data \
  --model_name="gpt-3.5-turbo" \
"""
import time
import json
import os
import random
import re
import string
from datetime import datetime

import numpy as np
import tqdm
import utils


def encode_prompt(query, prompt_papers):
    """Encode multiple prompt instructions into a single string."""
    prompt = open("relevancy_prompt.txt").read() + "\n"
    prompt += query['interest']

    for idx, task_dict in enumerate(prompt_papers):
        (title, authors, abstract) = task_dict["title"], task_dict["authors"], task_dict["abstract"]
        if not title:
            raise
        prompt += f"###\n"
        prompt += f"{idx + 1}. Title: {title}\n"
        prompt += f"{idx + 1}. Authors: {authors}\n"
        prompt += f"{idx + 1}. Abstract: {abstract}\n"
    prompt += f"\n Generate response:\n1."
    print(prompt)
    return prompt


def post_process_chat_gpt_response(paper_data, response, threshold_score=8):
    selected_data = []
    if response is None:
        return []
    json_items = response['message']['content'].replace("\n\n", "\n").split("\n")
    pattern = r"^\d+\. |\\"
    import pprint
    try:
        score_items = [
            json.loads(re.sub(pattern, "", line))
            for line in json_items if "relevancy score" in line.lower()]
    except Exception:
        pprint.pprint([re.sub(pattern, "", line) for line in json_items if "relevancy score" in line.lower()])
        raise RuntimeError("failed")
    pprint.pprint(score_items)
    scores = []
    for item in score_items:
        temp = item["Relevancy score"]
        if "/" in temp:
            scores.append(int(temp.split("/")[0]))
        else:
            scores.append(int(temp))
    if len(score_items) != len(paper_data):
        score_items = score_items[:len(paper_data)]
        hallucination = True
    else:
        hallucination = False

    for idx, inst in enumerate(score_items):
        # if the decoding stops due to length, the last example is likely truncated so we discard it
        if scores[idx] < threshold_score:
            continue
        output_str = "Title: " + paper_data[idx]["title"] + "\n"
        output_str += "Authors: " + paper_data[idx]["authors"] + "\n"
        output_str += "Link: " + paper_data[idx]["main_page"] + "\n"
        for key, value in inst.items():
            paper_data[idx][key] = value
            output_str += key + ": " + value + "\n"
        paper_data[idx]['summarized_text'] = output_str
        selected_data.append(paper_data[idx])
    return selected_data, hallucination


def find_word_in_string(w, s):
    return re.compile(r"\b({0})\b".format(w), flags=re.IGNORECASE).search(s)


def process_subject_fields(subjects):
    all_subjects = subjects.split(";")
    all_subjects = [s.split(" (")[0] for s in all_subjects]
    return all_subjects

def generate_relevance_score(
    all_papers,
    query,
    model_name="gpt-3.5-turbo",
    threshold_score=8,
    num_paper_in_prompt=4,
    temperature=0.4,
    top_p=1.0,
    sorting=True
):
    ans_data = []
    request_idx = 1
    hallucination = False
    for id in tqdm.tqdm(range(0, len(all_papers), num_paper_in_prompt)):
        prompt_papers = all_papers[id:id+num_paper_in_prompt]
        # only sampling from the seed tasks
        prompt = encode_prompt(query, prompt_papers)

        decoding_args = utils.OpenAIDecodingArguments(
            temperature=temperature,
            n=1,
            max_tokens=1072,  # hard-code to maximize the length. the requests will be automatically adjusted
            top_p=top_p,
        )
        request_start = time.time()
        response = utils.openai_completion(
            prompts=prompt,
            model_name=model_name,
            batch_size=1,
            decoding_args=decoding_args,
            logit_bias={"100257": -100},  # prevent the <|endoftext|> from being generated
            # "100265":-100, "100276":-100 for <|im_end|> and <endofprompt> token 
        )
        print ("response", response['message']['content'])
        request_duration = time.time() - request_start

        process_start = time.time()
        batch_data, hallu = post_process_chat_gpt_response(prompt_papers, response, threshold_score=threshold_score)
        hallucination = hallucination or hallu
        ans_data.extend(batch_data)

        print(f"Request {request_idx+1} took {request_duration:.2f}s")
        print(f"Post-processing took {time.time() - process_start:.2f}s")

    if sorting:
        ans_data = sorted(ans_data, key=lambda x: int(x["Relevancy score"]), reverse=True)
    
    return ans_data, hallucination

def run_all_day_paper(
    query={"interest":"", "subjects":["Computation and Language", "Artificial Intelligence"]},
    date=None,
    data_dir="../data",
    model_name="gpt-3.5-turbo",
    threshold_score=8,
    num_paper_in_prompt=8,
    temperature=0.4,
    top_p=1.0
):
    if date is None:
        date = datetime.today().strftime('%a, %d %b %y')
        # string format such as Wed, 10 May 23
    print ("the date for the arxiv data is: ", date)

    all_papers = [json.loads(l) for l in open(f"{data_dir}/{date}.jsonl", "r")]
    print (f"We found {len(all_papers)}.")

    all_papers_in_subjects = [
        t for t in all_papers
        if bool(set(process_subject_fields(t['subjects'])) & set(query['subjects']))
    ]
    print(f"After filtering subjects, we have {len(all_papers_in_subjects)} papers left.")
    ans_data = generate_relevance_score(all_papers_in_subjects, query, model_name, threshold_score, num_paper_in_prompt, temperature, top_p)
    utils.write_ans_to_file(ans_data, date, output_dir="../outputs")
    return ans_data


if __name__ == "__main__":
    query = {"interest":"""
    1. Large language model pretraining and finetunings
    2. Multimodal machine learning
    3. Do not care about specific application, for example, information extraction, summarization, etc.
    4. Not interested in paper focus on specific languages, e.g., Arabic, Chinese, etc.\n""",
    "subjects":["Computation and Language"]}
    ans_data = run_all_day_paper(query)