{
  "notes": "Training-ready GPIO-LLM dataset release built by scripts/build_release.py from the validated splits (data/splits), the prepared English corpus and configs/pretrain_mix.json. Stage 1 is the pretraining mixture (English + GPIO text); stage 2 is supervised fine-tuning on prompt/target pairs with a small English replay. Token counts are GPT-2 counts for web text and characters/4 elsewhere until a tokenizer is chosen.",
  "version": "gpio_llm_v1",
  "output_dir": "data/release/gpio_llm_v1",
  "manifest": "data/release/gpio_llm_v1.manifest.json",
  "card": "docs/DATASET_CARD.md",
  "seed": 20260914,
  "sft": {
    "splits_dir": "data/splits",
    "rows_per_shard": 100000,
    "eval_core": {
      "min_rows": 5000,
      "note": "Stratified subset of the full held-out eval: the same number of rows from every eval template (chosen by hashed id), so each unseen phrasing template counts equally. The full eval stays available."
    },
    "replay": {
      "component": "fineweb_edu_dedup",
      "token_share": 0.02,
      "note": "About 2% of fine-tuning tokens are general English documents from the pretraining corpus train split (docs/ENGLISH_DATASETS.md mixing table), to limit forgetting. Loss is on the whole document."
    },
    "disabled_components": {
      "massive_en_us": "Recommendation #2 (CC BY 4.0), 8-10% of fine-tuning once downloaded and mapped to the action schema.",
      "coconot": "Recommendation #3 (ODC-BY), about 5% once downloaded."
    }
  },
  "pretrain": {
    "mix": "configs/pretrain_mix.json",
    "docs_per_shard": 50000
  },
  "spec_targets": {
    "synthetic_examples": 100000,
    "eval_examples": 5000,
    "pretrain_tokens_min": 300000000,
    "source_text_bytes_min": 500000000
  },
  "accepted_shortfalls": {
    "gpio_examples": "85,118 GPIO examples accepted as sufficient by the project owner on 2026-09-14 (SPEC.md: start small; expand only when experiments show more data helps)."
  }
}