Datasets:
Tasks:
Text Generation
Modalities:
Text
Formats:
parquet
Languages:
Polish
Size:
10K - 100K
License:
File size: 4,360 Bytes
aae3233 | 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 | """Pack DynaWord-compatible CNC Parquet into Slayer uint16/doc-index files."""
from __future__ import annotations
import argparse
from array import array
import hashlib
import json
from pathlib import Path
import shutil
import struct
import sys
import pyarrow.parquet as pq
from tokenizers import Tokenizer
ROOT = Path(__file__).resolve().parents[1]
INPUT_DIRS = {"train": "data/train.parquet", "valid": "data/validation.parquet", "test": "data/test.parquet"}
OUTPUT_DIRS = {"train": "train", "valid": "valid", "test": "test"}
def sha256(path: Path) -> str:
digest = hashlib.sha256()
with path.open("rb") as handle:
for chunk in iter(lambda: handle.read(1 << 20), b""):
digest.update(chunk)
return digest.hexdigest()
def pack_split(input_root: Path, output_root: Path, split: str, tokenizer: Tokenizer) -> dict:
parquet_file = input_root / INPUT_DIRS[split]
if not parquet_file.is_file():
raise FileNotFoundError(f"missing Parquet file for {split}: {parquet_file}")
parquet_files = [parquet_file]
output_root.mkdir(parents=True, exist_ok=True)
bin_path = output_root / "train.bin"
if split != "train":
bin_path = output_root / f"{split}.bin"
idx_path = output_root / "docs.idx"
documents = 0
tokens = 0
with bin_path.open("wb") as bin_file, idx_path.open("wb") as idx_file:
offset = 0
for parquet_path in parquet_files:
table = pq.read_table(parquet_path, columns=["id", "text"])
for record_id, text in zip(table["id"].to_pylist(), table["text"].to_pylist()):
ids = tokenizer.encode(text).ids
if any(token_id > 65535 for token_id in ids):
raise ValueError(f"token id exceeds uint16 for {record_id}")
values = array("H", ids)
values.append(0)
if sys.byteorder != "little":
values.byteswap()
if values.itemsize != 2:
raise RuntimeError("uint16 packing requires two-byte unsigned array items")
bin_file.write(values.tobytes())
offset += len(values)
idx_file.write(struct.pack("<Q", offset))
documents += 1
tokens += len(values)
return {"documents": documents, "tokens_including_eos": tokens, "bin_sha256": sha256(bin_path), "idx_sha256": sha256(idx_path)}
def pack(input_root: Path, output_root: Path, tokenizer_path: Path, dataset_name: str) -> dict:
tokenizer = Tokenizer.from_file(str(tokenizer_path))
if tokenizer.token_to_id("<|endoftext|>") != 0:
raise ValueError("expected Slayer EOS token <|endoftext|> to have id 0")
output_root.mkdir(parents=True, exist_ok=True)
shutil.copy2(tokenizer_path, output_root / "tokenizer.json")
stats = {}
for split in INPUT_DIRS:
stats[split] = pack_split(input_root, output_root / OUTPUT_DIRS[split], split, tokenizer)
manifest = {
"dataset": dataset_name,
"version": json.loads((input_root / "manifest.json").read_text(encoding="utf-8")).get("version", "0.1.0"),
"format": "little-endian uint16 token ids; one EOS token 0 after every document",
"index": "little-endian uint64 document end offsets in tokens including EOS",
"tokenizer": "Slayer1B-1.06 tokenizer.json",
"tokenizer_sha256": sha256(tokenizer_path),
"vocab_size": tokenizer.get_vocab_size(),
"eos_id": 0,
"source": "../data/{train,validation,test}.parquet",
"splits": stats,
}
(output_root / "manifest.json").write_text(json.dumps(manifest, ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
return manifest
def main() -> None:
parser = argparse.ArgumentParser()
parser.add_argument("--dataset-name", default="CNC/CAM Slayer pretraining pack")
parser.add_argument("--input", type=Path, default=ROOT / "data" / "cnc_dynaword")
parser.add_argument("--output", type=Path, default=ROOT / "data" / "cnc_dynaword" / "slayer_v6")
parser.add_argument("--tokenizer", type=Path, default=ROOT / "models" / "Slayer1B-1.06" / "tokenizer.json")
args = parser.parse_args()
print(json.dumps(pack(args.input, args.output, args.tokenizer, args.dataset_name), ensure_ascii=False, indent=2))
if __name__ == "__main__":
main()
|