convert.py

from datasets import load_dataset, Features, Value, Sequence
from dataclasses import dataclass, field
import logging
from transformers import HfArgumentParser
from tqdm import tqdm
from typing import Dict, List
import json
import numpy as np
from itertools import islice

logger = logging.getLogger()
logger.setLevel(logging.INFO)
console_handler = logging.StreamHandler()
console_handler.setFormatter(
    logging.Formatter("[%(asctime)s %(levelname)s] %(message)s")
)
logger.handlers = [console_handler]


@dataclass
class ConversionAgruments:
    hardneg: str = field(metadata={"help": "Path to msmarco-hard-negatives.jsonl file"})
    out: str = field(metadata={"help": "Output path"})


@dataclass
class QRel:
    doc: int
    score: int


def load_msmarco(path: str, split) -> Dict[int, str]:
    dataset = load_dataset(path, split, split=split)
    cache: Dict[int, str] = {}
    for row in tqdm(dataset, desc=f"loading {path} split={split}"):
        index = int(row["_id"])
        cache[index] = row["text"]
    return cache


def load_qrel(path: str, split: str) -> Dict[int, List[QRel]]:
    dataset = load_dataset(path, split=split)
    print(dataset.features)
    cache: Dict[int, List[QRel]] = {}
    for row in tqdm(dataset, desc=f"loading {path} split={split}"):
        qid = int(row["query-id"])
        qrel = QRel(int(row["corpus-id"]), int(row["score"]))
        if qid in cache:
            cache[qid].append(qrel)
        else:
            cache[qid] = [qrel]
    return cache


def process_raw(
    qrels: Dict[int, List[QRel]],
    queries: Dict[int, str],
    corpus: Dict[int, str],
    hardneg: Dict[int, List[int]],
) -> List[Dict]:
    result = []
    for query, rels in tqdm(qrels.items(), desc="processing split"):
        pos = [
            {"doc": corpus[rel.doc], "score": rel.score}
            for rel in rels
            if rel.doc in corpus and rel.score > 0
        ]
        neg = [
            {"doc": corpus[doc], "score": 0.0}
            for doc in hardneg.get(query, [])
            if doc in corpus
        ]
        group = {"query": queries[query], "pos": pos, "neg": neg}
        result.append(group)
    return result


def load_hardneg(path: str):
    result: Dict[int, List[int]] = {}
    with open(path, "r") as jsonfile:
        for line in tqdm(jsonfile, total=808731, desc="loading hard negatives"):
            row = json.loads(line)
            scores: Dict[int, float] = {}
            for method, docs in row["neg"].items():
                for index, doc in enumerate(docs):
                    prev = scores.get(int(doc), 0.0)
                    scores[int(doc)] = prev + 1.0 / (60 + index)
            topneg = [
                doc
                for doc, score in sorted(
                    scores.items(), key=lambda x: x[1], reverse=True
                )
            ]
            result[int(row["qid"])] = topneg[:32]
    return result


def main():
    parser = HfArgumentParser((ConversionAgruments))
    (args,) = parser.parse_args_into_dataclasses()
    print(f"Args: {args}")
    hardneg = load_hardneg(args.hardneg)
    qrels = {
        "train": load_qrel("BeIR/msmarco-qrels", split="train"),
        "dev": load_qrel("BeIR/msmarco-qrels", split="validation"),
    }
    queries = load_msmarco("BeIR/msmarco", split="queries")
    corpus = load_msmarco("BeIR/msmarco", split="corpus")
    print("processing done")
    for split, data in qrels.items():
        dataset = process_raw(data, queries, corpus, hardneg)
        with open(f"{args.out}/{split}.jsonl", "w") as out:
            for item in dataset:
                json.dump(item, out)
                out.write("\n")
    print("done")


if __name__ == "__main__":
    main()