src/generator.py

import json
import re
import torch
from string import Template
from typing import List


def remove_citations(sent):
    return (
        re.sub(r"\[\d+", "", re.sub(r" \[\d+", "", sent))
        .replace(" |", "")
        .replace("]", "")
    )


class BaseGenerator:
    def __init__(self, prompt_path: str):
        self.prompt_path = prompt_path

        self.instruction, self.prompt_format = self._load_prompt(prompt_path)

    def _load_prompt(self, path: str):
        with open(path, "r") as f:
            prompt_config = json.load(f)

        instruction = prompt_config["instruction"]
        prompt_format = Template(prompt_config["prompt_format"])

        return instruction, prompt_format

    def _format_documents(self, doc_list: list) -> str:
        documents = ""
        for idx, doc in enumerate(doc_list):
            title = doc["title"]
            text = doc["text"]

            documents += f"Document [{idx + 1}](Title: {title}): {text}\n"
        return documents


class SentenceGenerator(BaseGenerator):
    def __init__(self, model, tokenizer, prompt_path: str):
        super().__init__(prompt_path)

        self.model = model
        self.tokenizer = tokenizer

    def format_prompt(self, question: str, doc_list: list) -> str:
        prompt = self.prompt_format.substitute(
            instruction=self.instruction,
            question=question,
            documents=self._format_documents(doc_list),
        )
        return prompt

    @torch.inference_mode()
    def generate(self, prompt: str) -> (str, List[int]):
        # * tokenization
        input_ids = self.tokenizer(prompt, return_tensors="pt")["input_ids"].to(
            self.model.device
        )
        prompt_len = input_ids.shape[1]
        # * generate until the end of the sentence
        outputs = self.model.generate(
            input_ids,
            max_new_tokens=128,
            eos_token_id=[self.tokenizer.eos_token_id, 29889],
        ).to("cpu")
        # * decode and remove the prompt
        sentence = self.tokenizer.decode(
            outputs[0][prompt_len:], skip_special_tokens=True
        ).strip()
        # * make sure that only one sentence is generated
        sentence = (
            sentence[: sentence.index("].") + 2] if "]." in sentence else sentence
        )
        # * remove citations
        sentence = remove_citations(sentence)

        return sentence, outputs[0]


class CitationGenerator(BaseGenerator):
    def __init__(self, model, tokenizer, prompt_path: str):
        super().__init__(prompt_path)

        self.model = model
        self.tokenizer = tokenizer

    def format_prompt(self, doc_list: list, sentence: str) -> str:
        prompt = self.prompt_format.substitute(
            instruction=self.instruction,
            documents=self._format_documents(doc_list),
            sentence=sentence,
        )
        return prompt

    @torch.inference_mode()
    def generate(self, prompt: str) -> str:
        # * tokenization
        input_ids = self.tokenizer(prompt, return_tensors="pt")["input_ids"].to(
            self.model.device
        )
        prompt_len = input_ids.shape[1]
        # * generate cited sentence
        outputs = self.model.generate(input_ids, max_new_tokens=128).to("cpu")
        # * decode and remove the prompt
        cited_sentence = self.tokenizer.decode(
            outputs[0][prompt_len:], skip_special_tokens=True
        ).strip()

        return cited_sentence

class QueryGenerator(BaseGenerator):
    def __init__(self, model, tokenizer, prompt_path: str):
        super().__init__(prompt_path)

        self.model = model
        self.tokenizer = tokenizer

    def format_prompt(self, question: str, context: str, claim: str, query_num: int) -> str:
        prompt = self.prompt_format.substitute(
            question=question,
            context=context,
            claim=claim,
            query_num=query_num
        )
        return prompt

    @torch.inference_mode()
    def generate(self, prompt: str) -> str:
        # * tokenization
        input_ids = self.tokenizer(prompt, return_tensors="pt")["input_ids"].to(
            self.model.device
        )
        prompt_len = input_ids.shape[1]
        # * generate query
        outputs = self.model.generate(input_ids, max_new_tokens=256).to("cpu")
        # * decode and remove the prompt
        query = self.tokenizer.decode(
            outputs[0][prompt_len:], skip_special_tokens=True
        ).strip()

        return query