File size: 7,194 Bytes
c7fac78
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
#!/usr/bin/env python3
"""Tokenize FineWeb-Edu parquet files to byte-level .bin format.

Spider-FLEXITOKENS uses byte-level vocab (272 tokens: 256 bytes + 16 specials).
This script converts raw text parquet files to uint16 .bin files ready for
fast loading by ByteLevelTokenizedDataset.

Output format:
  - tokens.uint16: concatenated byte values (0-255) + special tokens (256-271)
  - metadata.json: {num_tokens, num_samples, seq_len, etc.}
"""
import os
import sys
import json
import glob
import argparse
from pathlib import Path

import numpy as np
import pyarrow.parquet as pq
from tqdm import tqdm

# Sentinel tokens
BOS_ID = 257
EOS_ID = 258
PAD_ID = 256

def tokenize_text(text: str, max_bytes: int = 2048) -> list:
    """Convert text to byte-level token sequence with BOS/EOS."""
    byte_ids = list(text.encode('utf-8'))[:max_bytes]
    return [BOS_ID] + byte_ids + [EOS_ID]

def process_parquet(parquet_path: str, output_dir: str, seq_len: int = 2048, max_samples: int = 0):
    """Process a single parquet file to tokenized .bin."""
    print(f"Processing: {parquet_path}")
    output_path = os.path.join(output_dir, os.path.basename(parquet_path).replace('.parquet', '.bin'))
    
    tokens = []
    num_samples = 0
    
    try:
        pf = pq.ParquetFile(parquet_path)
        for batch in tqdm(pf.iter_batches(batch_size=1000, columns=["text"]), desc=os.path.basename(parquet_path)):
            texts = batch["text"].to_pylist()
            for text in texts:
                if not text:
                    continue
                token_ids = tokenize_text(text, max_bytes=seq_len - 2)  # -2 for BOS/EOS
                tokens.extend(token_ids)
                num_samples += 1
                if max_samples > 0 and num_samples >= max_samples:
                    break
            if max_samples > 0 and num_samples >= max_samples:
                break
    except Exception as e:
        print(f"Error processing {parquet_path}: {e}")
        return 0, 0
    
    if tokens:
        arr = np.array(tokens, dtype=np.uint16)
        arr.tofile(output_path)
        print(f"  Saved: {output_path} ({arr.nbytes / 1024**2:.1f} MB, {len(tokens):,} tokens)")
        return len(tokens), num_samples
    return 0, 0

def stream_from_huggingface(output_dir, seq_len=2048, num_shards=50, samples_per_shard=50000):
    """Stream FineWeb-Edu from HuggingFace and tokenize to .bin shards."""
    from datasets import load_dataset
    from tqdm import tqdm

    os.makedirs(output_dir, exist_ok=True)

    print(f"Streaming FineWeb-Edu from HuggingFace (10BT sample)")
    print(f"Output: {output_dir}")

    ds = load_dataset("HuggingFaceFW/fineweb-edu", "sample/10BT", split="train", streaming=True)

    total_tokens = 0
    total_samples = 0
    shard_tokens = []
    shard_count = 0

    for sample in tqdm(ds, desc="Tokenizing"):
        text = sample.get("text", "")
        if not text:
            continue
        token_ids = tokenize_text(text, max_bytes=seq_len - 2)
        shard_tokens.extend(token_ids)
        total_samples += 1

        if len(shard_tokens) >= samples_per_shard * seq_len:
            shard_count += 1
            arr = np.array(shard_tokens, dtype=np.uint16)
            shard_path = os.path.join(output_dir, f"shard_{shard_count:04d}.bin")
            arr.tofile(shard_path)
            print(f"  Shard {shard_count}: {shard_path} ({arr.nbytes / 1024**2:.1f} MB, {len(shard_tokens):,} tokens, {total_samples:,} samples)")
            total_tokens += len(shard_tokens)
            shard_tokens = []

        if shard_count >= num_shards:
            break

    if shard_tokens:
        shard_count += 1
        arr = np.array(shard_tokens, dtype=np.uint16)
        shard_path = os.path.join(output_dir, f"shard_{shard_count:04d}.bin")
        arr.tofile(shard_path)
        print(f"  Shard {shard_count}: {shard_path} ({arr.nbytes / 1024**2:.1f} MB, {len(shard_tokens):,} tokens)")
        total_tokens += len(shard_tokens)

    metadata = {
        "source": "HuggingFaceFW/fineweb-edu sample/10BT",
        "seq_len": seq_len,
        "vocab_size": 272,
        "dtype": "uint16",
        "total_samples": total_samples,
        "total_tokens": total_tokens,
        "num_shards": shard_count,
        "token_format": "BOS + UTF-8 bytes + EOS (byte-level, 272 vocab)",
    }
    with open(os.path.join(output_dir, "metadata.json"), 'w') as f:
        json.dump(metadata, f, indent=2)

    print(f"\nDone! Total: {total_samples:,} samples, {total_tokens:,} tokens, {shard_count} shards")
    print(f"Metadata saved to {output_dir}/metadata.json")


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("--input_dir", type=str, default="",
                        help="Input directory with parquet files (local mode)")
    parser.add_argument("--output_dir", type=str, default="/home/lamcodealong/fineweb_bytelevel",
                        help="Output directory for tokenized .bin files")
    parser.add_argument("--seq_len", type=int, default=2048,
                        help="Max sequence length (including BOS/EOS)")
    parser.add_argument("--max_samples", type=int, default=0,
                        help="Max samples per file (0=all, local mode only)")
    parser.add_argument("--max_files", type=int, default=0,
                        help="Max number of parquet files to process (0=all, local mode only)")
    parser.add_argument("--num_shards", type=int, default=50,
                        help="Number of output shards (HuggingFace mode)")
    parser.add_argument("--samples_per_shard", type=int, default=50000,
                        help="Approx samples per shard (HuggingFace mode)")
    parser.add_argument("--from_hf", action="store_true",
                        help="Stream from HuggingFace instead of local parquet files")
    args = parser.parse_args()

    if args.from_hf or not args.input_dir:
        stream_from_huggingface(args.output_dir, args.seq_len, args.num_shards, args.samples_per_shard)
        return

    os.makedirs(args.output_dir, exist_ok=True)

    parquet_files = sorted(glob.glob(os.path.join(args.input_dir, "*.parquet")))
    if args.max_files > 0:
        parquet_files = parquet_files[:args.max_files]

    print(f"Found {len(parquet_files)} parquet files")
    print(f"Output: {args.output_dir}")

    total_tokens = 0
    total_samples = 0

    for pq_file in parquet_files:
        n_tokens, n_samples = process_parquet(
            pq_file, args.output_dir, args.seq_len, args.max_samples
        )
        total_tokens += n_tokens
        total_samples += n_samples

    metadata = {
        "input_dir": args.input_dir,
        "seq_len": args.seq_len,
        "vocab_size": 272,
        "dtype": "uint16",
        "total_samples": total_samples,
        "total_tokens": total_tokens,
        "token_format": "BOS + UTF-8 bytes + EOS (byte-level, 272 vocab)",
    }
    with open(os.path.join(args.output_dir, "metadata.json"), 'w') as f:
        json.dump(metadata, f, indent=2)

    print(f"\nDone! Total: {total_samples:,} samples, {total_tokens:,} tokens")
    print(f"Metadata saved to {args.output_dir}/metadata.json")

if __name__ == "__main__":
    main()