RedPajamaV2/scripts/utils/get_counts.py

"""
Utility to get doc and character counts
"""

import os
import gzip
import orjson
import multiprocessing
from tqdm import tqdm
import json

def process_shard_unfiltered(
    shard
):
    global LANGUAGE
    global PARTITION
    global SIGNALS_DIR
    global SIGNALS_EXTENSION

    result = {
        "doc_count": 0,
        "char_count": 0
    }

    signals_file_path = os.path.join(
        SIGNALS_DIR,
        shard,
        f"{LANGUAGE}_{PARTITION}.{SIGNALS_EXTENSION}"
    )

    with gzip.open(signals_file_path, "r") as signals_file:
        for line in signals_file:
            signals_dict = orjson.loads(line)
            result["doc_count"] += 1
            result["char_count"] += signals_dict["quality_signals"]["ccnet_length"][0][2]

    return result

def process_shard_docids(
    shard
):
    global LANGUAGE
    global PARTITION
    global SIGNALS_DIR
    global SIGNALS_EXTENSION
    global DOCIDS_DIR

    docids_file_path = os.path.join(
        DOCIDS_DIR,
        shard,
        f"{LANGUAGE}_{PARTITION}.{DOCIDS_EXTENSION}"
    )
    docids = json.load(open(docids_file_path))
    docids_set = set(docids)

    result = {
        "doc_count": 0,
        "char_count": 0
    }

    signals_file_path = os.path.join(
        SIGNALS_DIR,
        shard,
        f"{LANGUAGE}_{PARTITION}.{SIGNALS_EXTENSION}"
    )

    with gzip.open(signals_file_path, "r") as signals_file:
        for line in signals_file:
            signals_dict = orjson.loads(line)
            if signals_dict["id"] in docids_set:
                result["doc_count"] += 1
                result["char_count"] += signals_dict["quality_signals"]["ccnet_length"][0][2]

    assert result["doc_count"] == len(docids_set), f"ERROR in counts for {shard=}"

    return result

def process_docids_store(
    processing_function
):
    with multiprocessing.Pool(NUM_CORES) as pool:
        shards_list = sorted(os.listdir(DOCIDS_DIR))
        all_results = list(tqdm(
            pool.imap(processing_function, shards_list),
            total=len(shards_list)
        ))
    all_shards_counts = {
        "doc_count": 0,
        "char_count": 0
    }
    for result in all_results:
        for k in list(all_shards_counts.keys()):
            all_shards_counts[k] += result[k]
    return all_shards_counts


if __name__ == "__main__":
    NUM_CORES = 50
    SNAPSHOT = "2023-14"
    LANGUAGE = "en"
    PARTITION = "head"
    DATA_ROOT_DIR = "/mnt/weka/peacock/enfm-dataprocessing/RedPajamaV2/data"
    SIGNALS_DIR = f"{DATA_ROOT_DIR}/quality_signals/{SNAPSHOT}"
    DOCIDS_DIR = f"{DATA_ROOT_DIR}/filtered_docids/quality_filtered_minus_clustered/{SNAPSHOT}"
    SIGNALS_EXTENSION = "signals.json.gz"
    DOCIDS_EXTENSION = "json"
    GET_COMPLETE_DATASET_COUNTS = False  # set True to count complete (unfiltered) dataset counts
                                        # but note that SIGNALS_DIR then must be that of unfiltered only
                                        # (not a subset, otherwise will get its counts)

    if GET_COMPLETE_DATASET_COUNTS:
        print(process_docids_store(
            process_shard_unfiltered
        ))
    else:
        print(process_docids_store(
            process_shard_docids
        ))

"""
Notes
- Counts format: (doc_count, char_count)
- RedPajamaV2 Snapshot 2023-14
    - unfiltered
        - head: (133M, 678.9B)
        - middle: (194.3M, 944.8B)
        - total: (327.3M, 1623.7B)
    - quality_filtered
        - head: (76.5M, 437.4B)
        - middle: (113.9M, 605.9B)
    - quality_filtered_minus_clustered
        - head: (67.6M, 382.8B)
        - middle: (102.8M, 539.5B)
        - total: (170.4M, 922.3B): 52% docs, 57% chars retained
"""