Datasets:
File size: 1,954 Bytes
a77eee2 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 | {
"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)."
}
}
|