jedld's picture
Upload folder using huggingface_hub
6e61f95 verified
|
Raw History Blame Contribute Delete
4.35 kB
metadata
license: other
tags:
  - mtg
  - magic-the-gathering
  - reinforcement-learning
  - distillation
  - imitation-learning
pretty_name: System One Choice — training data (IP-scubbed)

System One Choice — training datasets (copyright-scubbed)

Training data behind the play policy jedld/system-one-choice (model card documents the recipe: corpus LM warm-start → BC → PPO → hybrid soft-label distillation).

All third-party copyrightable text is omitted and replaced with pointers so downstream users regenerate it from the canonical upstream sources. What remains is generated by the open-source mtgcards engine (MIT) and its agents: engine game states, legal-action lists, decisions, outcomes, soft labels, and machine-generated strategist plans.

What was removed (and how to restore it)

Content Why Restore recipe
FineWeb crawled prose (8,000 docs) third-party web text fetch by doc id/url from HuggingFaceFW/fineweb sample-10BT (ODC-By-1.0)
Magic Comprehensive Rules text (106 chunks) © Wizards of the Coast fetch rules_url from the manifest
Card rules-text bodies everywhere (oracle/engine text inside card-detail blocks, corpus card docs) © Wizards of the Coast blocks keep the factual head (name, mana cost, type line, P/T, keywords) + a Scryfall lookup marker; regenerate verbatim via python -m mtgcards.rl.corpus build inside the MIT repo
Scryfall oracle JSON bodies (5,631 docs) © Wizards of the Coast name + set_code retained; re-fetch from Scryfall API

Machine-generated text (Strata/JEV-derived phase plans in plan/advice fields and in jev_state) is retained: it is synthetic model output, not third-party prose, and it is part of the hybrid-teacher training signal. Short functional action fragments inside options labels (e.g. mode 2: Sacrifice a token) are retained as they are the engine's legal-action identifiers.

Files

corpus/corpus.jsonl                      13,953 records: LM warm-start corpus with
                                         text omitted/replaced per source + pointers
corpus/…(see export_manifest.json)
trajectories/hybrid_bc.jsonl.gz          11,778 decision steps (69 games): BC corpus —
                                         hybrid Strata-plan + JEV Choice soft labels
                                         (probs), incl. paraphrase augmentation
trajectories/ppo_choice_fresh.jsonl.gz   109,152 decision steps (360 games): PPO
                                         self-play rollouts (league snapshots vs past
                                         checkpoints), reward = terminal ±1
export_manifest.json                     counts + scrub stats

Record fields (trajectories)

game, step, meta (stage/iter tags), player_idx, state_text (public-only engine view, card bodies replaced by lookup markers), options (legal-action labels), action (index taken), outcome (±1/0 seat outcome), logp, value, confidence, and — on teacher steps — probs / jev_probs (soft label distributions), jev_state, jev_conf, teacher, teacher_conf, plan, advice, paraphrased/paraphrase_tag.

Hidden information

Every state_text is the policy-visible surface the deciding seat is legitimately allowed to see: own hand, both battlefields, stack, public zone counts, previously-played cards. Opponent hand identities, library order, and face-down details are never present (enforced upstream by tests/test_hidden_info.py in the repo; scrubbing only ever removes text).

License

  • Engine-generated data (states, decisions, labels, plans): MIT, same as jedld/mtgcards.
  • Pointers/attribution only for: FineWeb (ODC-By-1.0) and Magic: The Gathering rules/card text (© Wizards of the Coast — see the repo NOTICE). This dataset redistributes none of it.

Citation

Cite the model repo jedld/system-one-choice and the engine per its CITATION.cff, and note that faithful reproduction of the original training inputs requires supplying the FineWeb and WotC sources yourself under their licenses.