{ "notes": "Token-weighted pretraining mixture read by scripts/pretrain_mix.py (iter_mixture). Weights are relative among enabled components; a component stops once it has contributed max_epochs passes over its train tokens, and the others fill the rest. Token counts are exact for the selected tokenizer (token_counts_dir, written by scripts/tokenizer_build.py count). Eval data never enters the mixture: gpio_actions is rendered from the train split only.", "target_tokens": 550000000, "target_note": "About 27 tokens per parameter for the recommended ~20M-parameter model; within the ~4 passes over repeated data that data-constrained scaling results show cost little (Muennighoff et al., 2023).", "seed": 20260914, "shuffle_buffer": 10000, "token_counts_dir": "data/tokenized/counts/gpio_llm_bpe_12k", "components": [ { "name": "fineweb_edu_dedup", "enabled": true, "kind": "manifest", "manifest": "data/sources/english/fineweb_edu_dedup.json", "weight": 0.6, "max_epochs": 3.5, "role": "General English (docs/ENGLISH_DATASETS.md #1). Up to 3.5 passes; data-constrained scaling results show little loss from up to about 4 epochs of repetition." }, { "name": "gpio_actions", "enabled": true, "kind": "jsonl", "render_from": "data/splits/train.jsonl", "files": { "train": [ "data/cleaned/english/gpio_actions/train.jsonl" ] }, "weight": 0.03, "max_epochs": 6, "role": "This project's validated GPIO examples as plain text (train split only), capped at 6 passes; the action behaviour itself is taught in fine-tuning." }, { "name": "tinystories_v2", "enabled": false, "kind": "manifest", "manifest": "data/sources/english/tinystories_v2.json", "weight": 0.12, "max_epochs": 1.0, "role": "Recommendation #4, not downloaded yet." }, { "name": "electronics_stackexchange", "enabled": false, "kind": "manifest", "manifest": "data/sources/english/electronics_stackexchange.json", "weight": 0.08, "max_epochs": 1.0, "role": "Recommendation #5, not downloaded yet (CC BY-SA)." } ] }