Download schema.json from ishagarg1103/gambit-zero: direct link, hf CLI and curl.
- Browser
- Download file 2.39 kB
-
https://huggingface.co/datasets/ishagarg1103/gambit-zero/resolve/main/schema.json
- Command line
-
hf download hf://datasets/ishagarg1103/gambit-zero/schema.json
-
curl -L -o schema.json https://huggingface.co/datasets/ishagarg1103/gambit-zero/resolve/main/schema.json
2.39 kB
| { | |
| "schema": "human_v3", | |
| "label_source": "human_game", | |
| "wdl_order": [ | |
| "win", | |
| "draw", | |
| "loss" | |
| ], | |
| "encoder": "pgx-v2-history8-ep-context8-v1", | |
| "game_parquet_schema": "game_id: string\nmetadata: string\nactions: list<item: uint16>\n child 0, item: uint16\n-- schema metadata --\nschema: 'human_v3'\nencoder: 'pgx-v2-history8-ep-context8-v1'\naction_map: 'pgx-2.6.0-chess-v2-uci-v1'\nrules: 'tinychess-auto-fide-pgx-v1'\nwdl_order: 'win,draw,loss'\nlegal_mask_bitorder: 'little'\nfeature_dtype: 'little_endian_float32'", | |
| "position_parquet_schema": "game_id: string\nsource_month: string\nsplit: string\nply: uint16\nhuman_move_uci: string\naction: uint16\noutcome_stm: int8\nwdl_class: uint8\nvalue_weight: float\npolicy_weight: float\ncolor: uint8\nfen: string\nrepetition_count: uint8\nposition_hash: fixed_size_binary[32]\ntraining_eligible: bool\nremaining_plies: uint16\nlabel_source: string\nplanes: fixed_size_binary[30720]\ncontext: fixed_size_binary[32]\nlegal_mask: fixed_size_binary[584]\n-- schema metadata --\nschema: 'human_v3'\nencoder: 'pgx-v2-history8-ep-context8-v1'\naction_map: 'pgx-2.6.0-chess-v2-uci-v1'\nrules: 'tinychess-auto-fide-pgx-v1'\nwdl_order: 'win,draw,loss'\nlegal_mask_bitorder: 'little'\nfeature_dtype: 'little_endian_float32'", | |
| "actions": "native Pgx uint16 index sequence in ply order", | |
| "feature_bytes": { | |
| "planes": 30720, | |
| "context": 32, | |
| "legal_mask": 584 | |
| }, | |
| "legal_mask": "4672 bits, little bit order, no padding", | |
| "value_weight": "0 for time forfeit, 1 otherwise; independent of win/draw/loss label", | |
| "policy_weight": 1.0, | |
| "provenance": "Per-game JSON metadata joined on unique game_id", | |
| "temporal_splits": { | |
| "train": "source months 2026-01 through 2026-07", | |
| "validation": "2026-08", | |
| "test": "2026-09" | |
| }, | |
| "horizon": "reject >512 plies and any replay truncation; never create a truncation draw", | |
| "early_environment_end": "reject source continuations after genuine automatic board termination", | |
| "full_legal_parity": "all accepted pre-move positions plus final board checked offline; mismatch aborts entire extraction", | |
| "raw_annotations": "never use [%eval] as a label", | |
| "storage_choice": "materialized_position_parquet primary for measured random access; retain compact_game_parquet for replay/provenance", | |
| "storage_measurement_report": "reports/m1_storage_benchmark.json" | |
| } | |