testbed/embeddings-benchmark__mteb/scripts/data/arxiv/script_clustering.py

from __future__ import annotations

import gzip
import os
from collections import Counter

import datasets
import jsonlines
import numpy as np
from tqdm import tqdm

np.random.seed(28042000)

d = datasets.load_dataset("mteb/raw_arxiv")["train"]
# d = d.select(range(1000))

main_categories = [
    "cs",
    "econ",
    "math",
    "eess",
    "q-bio",
    "q-fin",
    "stat",
    "astro-ph",
    "physics",
    "cond-mat",
    "quant-ph",
    "hep",
]


def cluster_stats(labels):
    (unique, counts) = np.unique(labels, return_counts=True)
    for u, c in zip(unique, counts):
        print(u, c)


def get_text(record, type="s2s"):
    if type == "s2s":
        return record["title"]
    elif type == "p2p":
        return record["title"] + " " + record["abstract"]
    raise ValueError


def match_main_category(category):
    for main_cat in main_categories:
        if main_cat in category:
            return main_cat
    return "other"


def main_category(category_string):
    category_list = category_string.split(" ")
    category_list = list(map(match_main_category, category_list))
    main_cat = Counter(category_list).most_common(1)[0][0]
    return main_cat


def sub_category(category_string):
    category_list = category_string.split(" ")
    return category_list[0]


main_cats = []
sub_cats = []
for paper in tqdm(d):
    main_cats.append(main_category(paper["categories"]))
    sub_cats.append(sub_category(paper["categories"]))

main_cats = np.array(main_cats)
sub_cats = np.array(sub_cats)

split_size = 25000
split_number = 10
indices = np.arange(len(main_cats))

splits = []

task_type = "p2p"

# Coarse splits
for k in tqdm(range(split_number)):
    np.random.shuffle(indices)
    current_indices = indices[:split_size]
    subset = d.select(current_indices)
    text = [get_text(item, task_type) for item in subset]
    labels = main_cats[current_indices].tolist()
    splits.append({"sentences": text, "labels": labels})

# Fine grained splits
for k in tqdm(range(split_number)):
    np.random.shuffle(indices)
    current_indices = indices[:split_size]
    subset = d.select(current_indices)
    text = [get_text(item, task_type) for item in subset]
    labels = sub_cats[current_indices].tolist()
    splits.append({"sentences": text, "labels": labels})

# Fine grained splits inside sub categories
for main_cat in tqdm(main_categories):
    indices = np.argwhere(main_cats == main_cat).flatten()
    if len(indices) == 0:
        print(f"No papers in {main_cat}")
        continue
    np.random.shuffle(indices)
    current_indices = indices[:split_size]
    subset = d.select(current_indices)
    text = [get_text(item, task_type) for item in subset]
    labels = sub_cats[current_indices].tolist()
    splits.append({"sentences": text, "labels": labels})

repository = "arxiv-clustering-p2p"
with jsonlines.open(f"{repository}/test.jsonl", "w") as f_out:
    f_out.write_all(splits)
# Compress
with open(f"{repository}/test.jsonl", "rb") as f_in:
    with gzip.open(f"{repository}/test.jsonl.gz", "wb") as f_out:
        f_out.writelines(f_in)
# Remove uncompressed file
os.remove(f"{repository}/test.jsonl")