paramgolf-mix-8b

An 8B-token, cross-source-deduplicated pretraining mixture for OpenAI's Parameter Golf tiny-LM challenge. Built to test whether adding code + math to a FineWeb-edu base moves Δbpb on the official FineWeb validation split at the ~30M-param / 8B-token regime.

Shards are in the trainer's native .bin format (256×int32 header [magic=20240520, version=1, num_tokens] + uint16 LE tokens, 10^8 tokens/shard), directly consumable by the reference train_gpt.py.

Mixing ratio

Source	Share	Tokens (SP1024)
FineWeb-edu (sample/10BT)	80%	6,400,000,000
StarCoder (python)	12%	960,000,000
OpenWebMath	8%	640,000,000
Total	100%	8,000,000,000

Ratio rationale: the eval is bpb on FineWeb web text, so the web source stays dominant; code/math are kept to a 20% band (math < code) because off-distribution syntax trades against web perplexity at tiny scale. See RESEARCH.md §5.

Source datasets (exact revisions)

Source	HF repo	Revision SHA	Config
FineWeb-edu	HuggingFaceFW/fineweb-edu	87f09149ef47	sample/10BT
StarCoder	bigcode/starcoderdata	9fc30b578ced	python
OpenWebMath	open-web-math/open-web-math	fde8ef8de230	data/train-00[0-3]*

Cross-source dedup (datatrove MinHash-LSH)

Applied across the merged union of all three sources (not per-source).

• Config: 5-grams, 14 buckets × 8 hashes (112 perm), 64-bit xxhash, seed 1, Jaccard ≈ 0.75 (matches FineWeb's published config).
• Input: 18,430,351 docs → 17,630,414 survivors (4.34% removed).
• Per-source dup rate: FineWeb 6.61%, OpenWebMath 2.51%, StarCoder 1.60%.
• Cross-source overlap: 7,315 clusters spanning ≥2 sources (24,780 docs); largest pair FineWeb&OpenWebMath (6,615 clusters).

Tokenizer

• SP1024 SentencePiece BPE, vocab 1024, copied (not retrained) from the reference repo (willdepueoai/parameter-golf :: datasets/tokenizers/fineweb_1024_bpe.model).
• SHA-256: 4f5e8adb109c66b4886963bc75a7befd73bda36d27fd7102df8e9e66503b0e2a
• Doc encoding: [bos_id=1] + sp.encode(text), no EOS (matches baseline APPEND_EOS=False).
• Round-trip: token-level idempotence encode(decode(encode(x)))==encode(x) holds 1200/1200 (100%). Strict decode(encode(x))==x is N/A — SP1024 normalizes whitespace by design, and the eval computes bpb from per-piece byte lengths, never reconstructing raw text.

Provenance

• Provenance hash (sha256 of provenance.json): cf0270b28efbd33cc917f9d1066370573a982d290fb63d37fc2c6253a024df5c
• Code git SHA: fbc5d91f834b27424e2c8fda2b5b26e067e26ab3 (branch m1-dataset-build)
• Build host: EC2 m7i.48xlarge, us-west-2. Seeds: normalize/minhash/pack = 1.
• Full provenance: provenance.json

Licenses (inherited from upstream)

This mixture redistributes derived tokens from three upstream corpora; their original licenses govern the corresponding content:

• FineWeb-edu — ODC-By 1.0 (HuggingFaceFW/fineweb-edu).
• StarCoder / The Stack — code under permissive/OSI licenses per the BigCode license stack; subject to the StarCoder data terms and opt-out.
• OpenWebMath — released under the terms on its dataset card; web content retains original authors' rights.

Use must comply with each upstream license. No validation-split documents are included (val is the frozen official Parameter Golf FineWeb split; never trained on).

Links

• Working repo / build code: https://github.com/neveram/parameter-golf (branch m1-dataset-build)
• Research notes: RESEARCH.md
• Challenge: https://github.com/openai/parameter-golf