File size: 9,948 Bytes
97ecad4 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 | """
Generic fixed-length character-level preprocessing for reasoning tasks.
This script builds the shared vocabulary, serializes each example as
[quiz_padded][response_padded], writes flattened uint16 train/val binaries, and
stores metadata needed by AR/Tom-CAT evaluation and training entrypoints.
"""
import argparse
import json
import os
import pickle
import random
import string
import numpy as np
def parse_args():
parser = argparse.ArgumentParser(description='Generic data preparation for fixed-length reasoning tasks')
# Input/output paths.
parser.add_argument('--data_path', type=str, default='data/3sat7_train.jsonl', help='Path to cd5 JSONL file')
parser.add_argument('--out_dir', type=str, default='data/sat/3sat7/k8', help='Output directory for processed data')
# Dataset schema.
parser.add_argument('--input_key', type=str, default='input', help='JSON key for the prompt/quiz')
parser.add_argument('--output_key', type=str, default='output', help='JSON key for the completion/response')
parser.add_argument('--meta_name', type=str, default='meta.pkl', help='Name of the output metadata file')
# Train/validation split.
parser.add_argument('--val_ratio', type=float, default=0.1, help='Ratio of data to use for validation (e.g., 0.1 for 10%)')
parser.add_argument('--seed', type=int, default=42, help='Random seed for shuffling and splitting')
# Vocabulary control.
parser.add_argument('--custom_vocab', type=str, default='', help='Comma-separated custom characters. If empty, auto-scans the dataset.')
return parser.parse_args()
def main():
args = parse_args()
os.makedirs(args.out_dir, exist_ok=True)
random.seed(args.seed)
# 1. Read the JSONL source file.
print(f"Loading data from {args.data_path}...")
data = []
with open(args.data_path, 'r', encoding='utf-8') as f:
for line in f:
if line.strip():
data.append(json.loads(line.strip()))
print(f"Loaded {len(data)} samples.")
# 2. Build the vocabulary.
special_tokens = ["<PAD>", "<SEP>", "<EOS>", "<MASK>", "$"]
# Global base character set shared across common reasoning tasks.
global_base_chars = [
"0", "1", "2", "3", "4", "5", "6", "7", "8", "9",
",", "+", "-", "/", "=", "*",
] + list(string.ascii_lowercase)
if args.custom_vocab:
print("Using custom vocabulary...")
base_chars = [c.strip() for c in args.custom_vocab.split(',') if c.strip()]
else:
print("Using Global Unified Vocabulary...")
base_chars = global_base_chars.copy()
# Scan the dataset to append any unseen characters beyond the shared base set.
all_chars = set()
for sample in data:
all_chars.update(list(str(sample.get(args.input_key, ''))))
all_chars.update(list(str(sample.get(args.output_key, ''))))
unseen_chars = set(all_chars) - set(base_chars) - set(special_tokens)
if unseen_chars:
print(f"⚠️ Notice: Found new characters not in global vocab: {unseen_chars}")
base_chars.extend(sorted(list(unseen_chars)))
# Keep special tokens unique and place them at the end of the vocabulary.
chars = [c for c in base_chars if c not in special_tokens] + special_tokens
stoi = {ch: i for i, ch in enumerate(chars)}
itos = {i: ch for i, ch in enumerate(chars)}
vocab_size = len(chars)
if vocab_size >= 65536:
raise ValueError(f"vocab_size={vocab_size} exceeds uint16 capacity")
print(f"Vocab size: {vocab_size}")
def encode(s):
return [stoi[c] for c in str(s)]
# 3. Compute the maximum raw input/output lengths.
max_quiz_len = 0
max_response_len = 0
for sample in data:
quiz = str(sample.get(args.input_key, ''))
response = str(sample.get(args.output_key, ''))
max_quiz_len = max(max_quiz_len, len(quiz))
max_response_len = max(max_response_len, len(response))
quiz_size = max_quiz_len + 1 # +1 for <SEP>
response_size = max_response_len + 1 # +1 for <EOS>
data_size = quiz_size + response_size
print(f"max_quiz_len={max_quiz_len}, max_response_len={max_response_len}")
print(f"quiz_size={quiz_size}, response_size={response_size}, data_size={data_size}")
# 4. Dedup by (input, output) FIRST, so the same problem can't land in both
# train and val (a content-level leak that inflates val/test metrics). Then
# shuffle (fixed seed) and split into disjoint train/validation sets.
_seen = set()
_deduped = []
for _s in data:
_key = (str(_s.get(args.input_key, '')), str(_s.get(args.output_key, '')))
if _key in _seen:
continue
_seen.add(_key)
_deduped.append(_s)
if len(_deduped) != len(data):
print(f"Dedup: {len(data)} -> {len(_deduped)} samples "
f"({len(data) - len(_deduped)} duplicates removed before split)")
data = _deduped
random.shuffle(data)
num_val = int(len(data) * args.val_ratio)
val_samples = data[:num_val]
train_samples = data[num_val:]
print(f"Split: {len(train_samples)} train samples, {len(val_samples)} val samples.")
# 5. Encode samples into the fixed-length serialization:
# [quiz chars + PAD ... + SEP][response chars + PAD ... + EOS]
def process_samples(samples, dataset_name):
processed_seqs = []
for idx, sample in enumerate(samples):
quiz = str(sample.get(args.input_key, ''))
response = str(sample.get(args.output_key, ''))
quiz_encoded = encode(quiz)
response_encoded = encode(response)
# Pad the quiz and response to their dataset-wide maximum lengths.
quiz_padded = quiz_encoded + [stoi["<PAD>"]] * (max_quiz_len - len(quiz_encoded)) + [stoi["<SEP>"]]
response_padded = response_encoded + [stoi["<PAD>"]] * (max_response_len - len(response_encoded)) + [stoi["<EOS>"]]
seq = quiz_padded + response_padded
if len(seq) != data_size:
print(f"[{dataset_name}] Skipping invalid sequence at index {idx}: seq_len={len(seq)}, expected={data_size}")
continue
processed_seqs.extend(seq)
return processed_seqs
train_data = process_samples(train_samples, "Train")
val_data = process_samples(val_samples, "Val")
print(f"Raw train tokens: {len(train_data)}")
print(f"Raw val tokens: {len(val_data)}")
# Print a few decoded examples to verify the packing protocol.
print("\n" + "="*60)
print("VERIFICATION: Checking a few processed training samples...")
print("="*60)
num_examples_to_print = 3
if len(train_data) >= data_size * num_examples_to_print:
for i in range(num_examples_to_print):
start_idx = i * data_size
end_idx = start_idx + data_size
sample_seq = train_data[start_idx:end_idx]
decoded_seq = [itos[token_id] for token_id in sample_seq]
quiz_part = "".join(decoded_seq[:quiz_size])
resp_part = "".join(decoded_seq[quiz_size:])
print(f"--- Example {i+1} ---")
print(f"Padded Quiz (len={len(quiz_part)}): {quiz_part}")
print(f"Padded Response (len={len(resp_part)}): {resp_part}")
print()
else:
print("Not enough data to print examples.")
print("="*60 + "\n")
# 6. Truncate any trailing partial example if earlier skips broke alignment.
def truncate_to_block(data_list, block_size, name):
remainder = len(data_list) % block_size
if remainder != 0:
print(f"Truncating {name} data by {remainder} tokens to align with block size.")
return data_list[:-remainder]
return data_list
train_data = truncate_to_block(train_data, data_size, "train")
val_data = truncate_to_block(val_data, data_size, "val")
# 7. Convert to uint16 arrays and sanity-check the vocabulary range.
train_bin = np.array(train_data, dtype=np.uint16)
val_bin = np.array(val_data, dtype=np.uint16)
assert train_bin.max() < vocab_size, f"Dirty data detected! Max token {train_bin.max()} >= vocab_size {vocab_size}"
if len(val_bin) > 0:
assert val_bin.max() < vocab_size, f"Dirty data detected! Max token {val_bin.max()} >= vocab_size {vocab_size}"
# Save flattened binary files.
train_bin.tofile(os.path.join(args.out_dir, 'train.bin'))
val_bin.tofile(os.path.join(args.out_dir, 'val.bin'))
# 8. Save metadata describing the serialization protocol.
meta = {
'format_version': 'fixed_length_char_v1',
'vocab_size': vocab_size,
'stoi': stoi,
'itos': itos,
'block_size': data_size - 1,
'quiz_size': quiz_size,
'response_size': response_size,
'data_size': data_size,
'max_quiz_len': max_quiz_len,
'max_response_len': max_response_len,
'max_input_len': max_quiz_len,
'max_output_len': max_response_len,
'input_key': args.input_key,
'output_key': args.output_key,
'special_tokens': special_tokens,
'pad_token': '<PAD>',
'sep_token': '<SEP>',
'eos_token': '<EOS>',
'mask_token': '<MASK>',
'dollar_token': '$',
'tokenizer_type': 'char',
'serialization': 'quiz_pad_sep + response_pad_eos',
'dtype': 'uint16',
'data_path': args.data_path,
'val_ratio': args.val_ratio,
'seed': args.seed,
}
meta_path = os.path.join(args.out_dir, args.meta_name)
with open(meta_path, 'wb') as f:
pickle.dump(meta, f)
print(f"✅ Data successfully prepared in '{args.out_dir}'.")
print(f" Saved train.bin, val.bin, and {args.meta_name}.")
if __name__ == "__main__":
main()
|