DALabCommunity/Haidass-translate-Datasets

Haidass Translate — V3.4 Stage1 full dataset

This is the complete frozen V3.4 Stage1 training input, not a sample. 中文说明、来源配比与审计边界见 DATA_RECIPE.md。

View	Directional examples	Purpose
data/train-*.jsonl.gz	18,935,335	Exact final eight-column input used for training
provenance/unique-*.jsonl.gz	16,579,656	Unique corpus under the original normalized exact-dedup policy, with richer provenance
provenance/replay-*.jsonl.gz	2,355,765	Intentional Chinese→English replay, not additional unique data

The provenance views reconstruct the original pre-tail-filter exposure stream (18,935,421 rows). Do not append them to the default training split: that would count the same examples again. The final training view excludes 86 overlength examples whose packed targets lost their end marker. Those exclusions and historical audits are retained under audit/.

Load

from datasets import load_dataset
ds = load_dataset("DALabCommunity/Haidass-translate-Datasets", split="train", streaming=True)
example = next(iter(ds))

Columns: origin_id, src (corpus label), direction (en-zh / zh-en), instruction, input, output, provenance, pair_group. Examples are directional supervised records, not bilingual pair counts. No extra generation, cleaning, sampling, reserialization or row reordering was performed for this publication. Gzip shards are lossless byte slices of the frozen JSONL streams, 250,000 LF-delimited records each except the last shard of each view.

Verify and restore

python scripts/verify_restore.py --root /path/to/downloaded/repo
# Optional: restore all four original JSONL views into a NEW directory.
python scripts/verify_restore.py --root /path/to/downloaded/repo --restore-to /path/to/new-directory

PACKAGE_MANIFEST.json contains per-shard compressed and raw SHA-256, byte sizes, row counts, complete-stream hashes and the full-training direction/source census. SHA256SUMS covers the compressed data shards and the exclusion ledger. FILES_SHA256SUMS covers every published file except itself and the Git-managed .gitattributes. No training weights, credentials, machine-specific caches or packed tensor binaries are included; the complete text training inputs, provenance and packing audit are included.

Historical scope and limitations

V3.4 Stage1 was trained independently from the original Haidass base; it did not continue V3.3 RC1, the old 70/30 fusion, or the current RL-balanced release. Publishing this dataset does not claim it was the direct training corpus of those released weights or that V3.4 was universally superior on held-out evaluation.

The recipe combines inherited cleaned parallel corpora, validated HY-MT1.5-7B Chinese→English retranslations and deliberate high-quality replay. Known pair-level exclusions and exact cross-pool duplicates were removed; a fresh exhaustive cross-pool MinHash pass was not completed. Historical decontamination checks do not establish absence of all semantic leakage. Legacy web/subtitle noise can remain; no claim of zero translation errors is made. The near-50/50 target-token balance describes the pre-tail-filter exposure stream, not equal numbers of directional rows. See DATA_RECIPE.md and LICENSE_NOTICE.md before use.