jedld commited on
Commit
6e61f95
·
verified ·
1 Parent(s): ea02f86

Upload folder using huggingface_hub

Browse files
.gitattributes CHANGED
@@ -58,3 +58,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
58
  # Video files - compressed
59
  *.mp4 filter=lfs diff=lfs merge=lfs -text
60
  *.webm filter=lfs diff=lfs merge=lfs -text
 
 
58
  # Video files - compressed
59
  *.mp4 filter=lfs diff=lfs merge=lfs -text
60
  *.webm filter=lfs diff=lfs merge=lfs -text
61
+ corpus/corpus.jsonl filter=lfs diff=lfs merge=lfs -text
README.md ADDED
@@ -0,0 +1,88 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: other
3
+ tags:
4
+ - mtg
5
+ - magic-the-gathering
6
+ - reinforcement-learning
7
+ - distillation
8
+ - imitation-learning
9
+ pretty_name: System One Choice — training data (IP-scubbed)
10
+ ---
11
+
12
+ # System One Choice — training datasets (copyright-scubbed)
13
+
14
+ Training data behind the play policy
15
+ **[jedld/system-one-choice](https://huggingface.co/jedld/system-one-choice)**
16
+ (model card documents the recipe: corpus LM warm-start → BC → PPO →
17
+ hybrid soft-label distillation).
18
+
19
+ **All third-party copyrightable text is omitted and replaced with pointers**
20
+ so downstream users regenerate it from the canonical upstream sources.
21
+ What remains is generated by the open-source
22
+ [mtgcards engine](https://github.com/jedld/mtgcards) (MIT) and its agents:
23
+ engine game states, legal-action lists, decisions, outcomes, soft labels,
24
+ and machine-generated strategist plans.
25
+
26
+ ## What was removed (and how to restore it)
27
+
28
+ | Content | Why | Restore recipe |
29
+ | --- | --- | --- |
30
+ | FineWeb crawled prose (8,000 docs) | third-party web text | fetch by doc `id`/`url` from [HuggingFaceFW/fineweb](https://huggingface.co/datasets/HuggingFaceFW/fineweb) `sample-10BT` (ODC-By-1.0) |
31
+ | Magic Comprehensive Rules text (106 chunks) | © Wizards of the Coast | fetch `rules_url` from the manifest |
32
+ | 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 |
33
+ | Scryfall oracle JSON bodies (5,631 docs) | © Wizards of the Coast | name + `set_code` retained; re-fetch from Scryfall API |
34
+
35
+ Machine-generated text (Strata/JEV-derived phase plans in `plan`/`advice`
36
+ fields and in `jev_state`) is retained: it is synthetic model output, not
37
+ third-party prose, and it is part of the hybrid-teacher training signal.
38
+ Short functional action fragments inside `options` labels (e.g.
39
+ `mode 2: Sacrifice a token`) are retained as they are the engine's
40
+ legal-action identifiers.
41
+
42
+ ## Files
43
+
44
+ ```
45
+ corpus/corpus.jsonl 13,953 records: LM warm-start corpus with
46
+ text omitted/replaced per source + pointers
47
+ corpus/…(see export_manifest.json)
48
+ trajectories/hybrid_bc.jsonl.gz 11,778 decision steps (69 games): BC corpus —
49
+ hybrid Strata-plan + JEV Choice soft labels
50
+ (probs), incl. paraphrase augmentation
51
+ trajectories/ppo_choice_fresh.jsonl.gz 109,152 decision steps (360 games): PPO
52
+ self-play rollouts (league snapshots vs past
53
+ checkpoints), reward = terminal ±1
54
+ export_manifest.json counts + scrub stats
55
+ ```
56
+
57
+ ### Record fields (trajectories)
58
+
59
+ `game`, `step`, `meta` (stage/iter tags), `player_idx`, `state_text`
60
+ (public-only engine view, card bodies replaced by lookup markers),
61
+ `options` (legal-action labels), `action` (index taken), `outcome`
62
+ (±1/0 seat outcome), `logp`, `value`, `confidence`, and — on teacher
63
+ steps — `probs` / `jev_probs` (soft label distributions), `jev_state`,
64
+ `jev_conf`, `teacher`, `teacher_conf`, `plan`, `advice`,
65
+ `paraphrased`/`paraphrase_tag`.
66
+
67
+ ## Hidden information
68
+
69
+ Every `state_text` is the policy-visible surface the deciding seat is
70
+ legitimately allowed to see: own hand, both battlefields, stack, public
71
+ zone counts, previously-played cards. Opponent hand identities, library
72
+ order, and face-down details are never present (enforced upstream by
73
+ `tests/test_hidden_info.py` in the repo; scrubbing only ever removes text).
74
+
75
+ ## License
76
+
77
+ - Engine-generated data (states, decisions, labels, plans): MIT, same as
78
+ [jedld/mtgcards](https://github.com/jedld/mtgcards).
79
+ - Pointers/attribution only for: FineWeb (ODC-By-1.0) and Magic:
80
+ The Gathering rules/card text (© Wizards of the Coast — see the repo
81
+ `NOTICE`). This dataset redistributes none of it.
82
+
83
+ ## Citation
84
+
85
+ Cite the model repo [jedld/system-one-choice](https://huggingface.co/jedld/system-one-choice)
86
+ and the engine per its `CITATION.cff`, and note that faithful reproduction
87
+ of the original training inputs requires supplying the FineWeb and
88
+ WotC sources yourself under their licenses.
corpus/corpus.jsonl ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:92bfcf5641d938779c701be47bf53a389550f35154ddd227b5f2c37b4dea5c31
3
+ size 36679187
corpus/manifest.json ADDED
@@ -0,0 +1,53 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "built_at": "2026-11-20T04:24:46.958776+00:00",
3
+ "out_dir": "data/systemone_corpus",
4
+ "doc_count": 13953,
5
+ "total_chars": 58295802,
6
+ "by_source": {
7
+ "fineweb": 8000,
8
+ "card_catalog": 168,
9
+ "oracle_json": 5631,
10
+ "gamelog": 48,
11
+ "mtg_rules": 1,
12
+ "mtg_rules_chunk": 105
13
+ },
14
+ "counts": {
15
+ "fineweb": 8000,
16
+ "cards": 5799,
17
+ "gamelogs": 48,
18
+ "rules": 106
19
+ },
20
+ "sources": {
21
+ "fineweb": {
22
+ "repo": "HuggingFaceFW/fineweb",
23
+ "config": "sample-10BT",
24
+ "docs": 8000,
25
+ "seed": 42,
26
+ "note": "Bounded stream prefix / reservoir — not representative of all FineWeb.",
27
+ "license": "ODC-By-1.0"
28
+ },
29
+ "cards": {
30
+ "catalog": "mtgcards.card_catalog + cardref.describe_card",
31
+ "oracle_dirs": [
32
+ "data/standard/cards",
33
+ "data/cards"
34
+ ]
35
+ },
36
+ "gamelogs": {
37
+ "dir": "data/game_logs"
38
+ },
39
+ "rules": {
40
+ "url": "https://media.wizards.com/2025/downloads/MagicCompRules%2020250606.txt",
41
+ "path": "data/rules/comprehensive_rules.txt",
42
+ "attribution": "Wizards of the Coast — Magic Comprehensive Rules"
43
+ }
44
+ },
45
+ "files": {
46
+ "corpus": "corpus.jsonl",
47
+ "fineweb": "sources/fineweb.jsonl",
48
+ "cards": "sources/cards.jsonl",
49
+ "gamelogs": "sources/gamelogs.jsonl",
50
+ "rules": "sources/rules.jsonl",
51
+ "rules_txt": "sources/rules.txt"
52
+ }
53
+ }
export_manifest.json ADDED
@@ -0,0 +1,82 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "generated_at": "2026-11-20T20:16:29",
3
+ "scrub": {
4
+ "cards_mapped": 140,
5
+ "unresolved_names": [
6
+ "\"3/1996.",
7
+ "\"Challenges for Indian Companies in the Financial Services KPO Business\"",
8
+ "\"European Union Monitoring Mission",
9
+ "\"How does KPO differ from BPO\"",
10
+ "\"International Security Assistance Force",
11
+ "\"Mumbai v Australians at Mumbai",
12
+ "\"Padma Awards Directory",
13
+ "\"Population by sex and age, by settlements, census 2001",
14
+ "\"SEIS Journal of management\"",
15
+ "\"The Philippines Animation Industry\"",
16
+ "\"send new email\"",
17
+ "$ 997 bulk discount plan",
18
+ "$800,000 in road and parking lot maintenance",
19
+ "- All footwear: very small amount",
20
+ "... The composition and distribution of polycyclic aromatic hydrocarbons",
21
+ "...of Intellect",
22
+ "...of Power",
23
+ "...of Spirit",
24
+ "...of Stamina",
25
+ "...of the Bear",
26
+ "...of the Boar",
27
+ "...of the Eagle",
28
+ "...of the Monkey",
29
+ "...of the Owl",
30
+ "...of the Whale",
31
+ "...of the Wolf",
32
+ "03/05/2008 : Mimirara Mirai",
33
+ "04/04/2012 : Yuke Yuke Go Go Onna no Ko!!",
34
+ "04/04/2013 : Tickeee Daisakusen! ~Yavay~ / Hyper Tickeee Queen no Uta",
35
+ "04/07/2010 : 444 no Kiseki",
36
+ "05/08/2007 : Hyper Yo Ban 3 ~Natsu no Shou~",
37
+ "07/01/2007 : Zoku Hyper Yo Ban",
38
+ "1 - 2 cups",
39
+ "1 can",
40
+ "1 cup",
41
+ "1 lb",
42
+ "1 letter of Phebe Scattergood, 6 letters",
43
+ "1 letter of Thomas Scattergood",
44
+ "1 tablespoon olive oil",
45
+ "1/3 Prankster",
46
+ "1/8 teaspoon salt",
47
+ "10 Easy Science Experiments For Kids 1. Bouncy Rubber Egg Experiment",
48
+ "10 Triple Jeopardy",
49
+ "10 Weeks or longer",
50
+ "10/07/2013 : Hanabeat",
51
+ "10/12/2014 : Bukubuku / xYOxx",
52
+ "10min \u2013 poster presentation",
53
+ "11 The Last of the Mohawks",
54
+ "11/11/2009 : Oh! Seishun!!",
55
+ "12.5ft x 17ft Trampoline"
56
+ ]
57
+ },
58
+ "stats": {
59
+ "corpus": {
60
+ "docs": 13953,
61
+ "by_source": {
62
+ "fineweb": 8000,
63
+ "card_catalog": 168,
64
+ "oracle_json": 5631,
65
+ "gamelog": 48,
66
+ "mtg_rules": 1,
67
+ "mtg_rules_chunk": 105
68
+ },
69
+ "residual_lines": 0
70
+ },
71
+ "hybrid_bc": {
72
+ "games": 69,
73
+ "steps": 11778,
74
+ "residual_lines": 0
75
+ },
76
+ "ppo_selfplay": {
77
+ "games": 360,
78
+ "steps": 109152,
79
+ "residual_lines": 0
80
+ }
81
+ }
82
+ }
trajectories/hybrid_bc.jsonl.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f77f0abdaca3c2c39192bec08a9db8005cb2ca50e3fe987d04aba9b6316774fa
3
+ size 3365479
trajectories/ppo_choice_fresh.jsonl.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1dd47554322064f26d729108c71c4ba5b435d7a5a03f303cc03b58d98a5a9a6b
3
+ size 18834208