nht10/cx_sampled_old

Older CulturaX sampled pool used for the B200 experiments

This release preserves the original text and document order of the local prepare_data.py sampling run. It is separate from nht10/cx_sampled (the newer ten-language 30B/10B pool). Use UPLOAD_COMPLETE.json and pin the completed commit.

Source: nguyenhuuthuat09/CulturaX_sampled, revision b19d850278693d37113c197857cc6328fa5c6881. Tokenizer: Qwen/Qwen3-0.6B-Base, revision da87bfb608c14b7cf20ba1ce41287e8de496c0cd. Sampling seed: 42. Text is unpacked and unmodified.

Sampling targets: 30B English tokens, 1B per other language, followed by roughly 10M validation tokens per language. These are original budget targets counted without EOS, with whole-document rounding; exact token totals and per-document token indexes were not saved by this older sampler. Training appends an explicit <|endoftext|> at every document boundary before packing. Other tokenizers can use the text, but the original token budgets will not apply.

Every split's ordered text SHA256 was recomputed and matched to the original local sampling receipt before upload. All six language/split document counts and training shard counts match the B200 sampling receipt. Full cross-machine B200 text-hash equality remains unverified because its receipt has no text hashes.

Subset	Training documents	Validation documents
qwen3_0.6b_base_en_30B	36,595,514	11,822
qwen3_0.6b_base_vi_1B	927,135	9,251
qwen3_0.6b_base_zh_1B	964,420	9,802
qwen3_0.6b_base_ru_1B	805,963	7,963
qwen3_0.6b_base_de_1B	974,614	9,596
qwen3_0.6b_base_ar_1B	846,723	8,337

from datasets import load_dataset
data = load_dataset("nht10/cx_sampled_old", "qwen3_0.6b_base_en_30B", revision="<completed-commit>")

The Arrow folders also support load_from_disk: concatenate train/shard_* in sorted order; validation/shard_000000 holds the original validation split. For B200's original directory layout, place the manifest-ordered train shards under Qwen_Qwen3-0.6B-Base/train// and the validation shard contents directly under Qwen_Qwen3-0.6B-Base/eval//. Renaming folders does not change their rows or order. Each shard needs its state.json, dataset_info.json and all Arrow files listed in state.json. No training caches are included.

The sampler selects train documents first, then validation positions. It has no additional text-hash or near-duplicate deduplication, so duplicate text may occur across splits. No stronger disjointness claim is made. See frozen implementation, selected raw-file manifest and receipts in provenance. Source data terms apply; no new ownership or license to the underlying web documents is claimed.