jvonrad's picture
Repo renamed: update names, link the collection
8fc3d36 verified
|
Raw History Blame Contribute Delete
3.45 kB
metadata
tags:
  - pretraining
  - tokenized
  - xscript-pretraining

XScript-Pretraining packed training shards (EN, AR)

Part of the Alignment without Transfer collection — data for the paper Alignment without Transfer: Tokenization Impacts Cross-lingual Representations during Multilingual Language Model Pretraining (forthcoming). Checkpoints: jvonrad/alignment-without-transfer.

The packed token shards (shards/*/pool_*.bin, 1,248 files, ~556 GB) were removed on 2026-11-18. They were only needed for the abandoned EN-AR 100B-token extension runs. Everything needed to rebuild or verify them is kept: the source file list and byte budget per pool (pools/*/stats.json), the tokenizers (tokenizers/), per-shard token counts (shards/*/index.json, meta.json) and sha256 hashes of every removed shard (shards/shards_manifest.json). Rebuild by re-pooling the listed FineWeb-HQ / FineWeb2-HQ parquet files and packing with xscript pack (see the GitHub repo); the holdout text (holdout/) is kept as-is because holdout BPB is reported on it. The shard files remain in this repo's git history until it is squashed.

Pre-tokenized training data for the XScript-Pretraining project (https://github.com/jvonrad/XScript-Pretraining): English and Arabic pools from FineWeb-HQ / FineWeb2-HQ, each packed with both of the project's main tokenizers. Uploaded so the EN-AR runs (including the 100B-token extension) could be trained on a machine without the original cluster's scratch storage.

Layout

shards/<lang>__<tok>/pool_NNNNN.bin   uint16 token ids, documents packed as <bos> doc <eos> <bos> doc <eos> ...
shards/<lang>__<tok>/index.json       {"pool_NNNNN.bin": n_tokens, ...} for every shard
shards/<lang>__<tok>/meta.json        lang, tokenizer, total_tokens, n_shards
shards/shards_manifest.json           sha256 + byte size of every shard file
tokenizers/unigram_{destarved,starved}/{sp.model,meta.json}
pools/<lang>/stats.json               pool construction stats: byte budget, docs, source parquet files, holdout file
holdout/<lang>_00000.jsonl.zst        the in-domain holdout text (first ~30 MB of the first source parquet, excluded from the pool)
code/xscript-pretraining-<sha>.bundle git bundles of the code at upload time (47d91c5, f6d370a)
shard dir tokenizer total tokens shards
en__unigram_destarved fair (5 languages) 64.40B 266
en__unigram_starved starved (~419 languages) 75.89B 266
ar__unigram_destarved fair (5 languages) 55.44B 358
ar__unigram_starved starved (~419 languages) 82.44B 358

About 556 GB in total. Sources: epfml/FineWeb-HQ (en), epfml/FineWeb2-HQ arb_Arab (ar). Each .bin is one pool shard, shuffled within the shard with a fixed seed; vocab size is 65,536 so uint16 is exact. The format is the one written by xscript pack (src/xscript/data/pack.py) and read by the deterministic loader src/xscript/data/loader.py in the GitHub repo, which memory-maps the .bin files listed in index.json. Verify a download against shards_manifest.json (or with scripts/_verify_shards.py) before training.

Do not rename files: the loader resolves shards through index.json.