ricoz/tongits-benchmark

Tongits Offline RL & Imitation Learning Benchmark Dataset

Curated benchmark dataset of 10,000 full-game simulation trajectories (366,479 decision steps) across 3-player expert Tongits self-play from Tongits.io.

Dataset Summary

• Episodes: 10,000 completed hands.
• Decision Steps: 366,479 individual decision steps.
• Observation Features: 349-dimensional float state space encoded as sparse CompactObservation objects.
• Action Space: 75 discrete action choices with sparse legal action index masks.
• Payoffs: Settled monetary units per player reflecting zero-sum terminal rules and dues.

File Structure

• manifest.json: Metadata tying the run together (seeds, rules version, layout sizes, dues multipliers).
• shard-000.jsonl.gz to shard-009.jsonl.gz: 10 gzipped JSONL shards, each containing ~1,000 deterministic episodes.

Schema Specification

Each line in a .jsonl.gz shard is a JSON-encoded TrajectoryEpisode dictionary matching packages/engine/src/rl-trajectory.ts:

{
  "seed": "r5-bc-0",
  "players": ["a", "b", "c"],
  "rulesVersion": 3,
  "actors": ["league", "league", "league"],
  "steps": [
    {
      "seat": 0,
      "obs": {
        "hand": [1, 2, 3, 4, 9, 12, 20, 21, 27, 31, 44, 46, 51],
        "top": -1,
        "discards": [],
        "melds": [[], [], []],
        "scalars": [0.3, 0.8667, 0.8, 0.8, ...]
      },
      "legal": [7, 8, 9, 10, 15, 18, 26, 27, 33, 37, 50, 52, 57, 58],
      "action": 58
    }
  ],
  "rewards": [4.0, -2.0, -2.0],
  "reason": "TONGITS",
  "carryIn": null
}

Fields:

• seed: String deal seed guaranteeing byte-for-byte reproducibility.
• players: Array of seat IDs (["a", "b", "c"]).
• rulesVersion: Engine rules version integer (3).
• actors: Policy specification per seat (["league", "league", "league"]).
• steps: Sequential policy decisions taken during the hand:
  • seat: Acting seat index into players (0, 1, or 2).
  • obs: CompactObservation sparse object (hand, top, discards, melds, scalars), expanding to the dense 349-dimensional float state vector via expandObservation.
  • legal: Legal action indices (sparse form of the 75-dim boolean action mask).
  • action: Integer index of the discrete action taken (0..74).
• rewards: Per-seat terminal monetary outcomes in fee units, indexed like players ([4.0, -2.0, -2.0]).
• reason: Terminal hand outcome ("TONGITS", "FIGHT", "DECK_EMPTY").
• carryIn: Optional deal-time carryover Match Pool state ({"streakSeat": null, "poolUnits": 0}).

Quickstart: Python

import gzip
import json
from huggingface_hub import hf_hub_download

# Download a shard
file_path = hf_hub_download(
    repo_id="ricoz/tongits-benchmark",
    filename="shard-000.jsonl.gz",
    repo_type="dataset"
)

# Iterate over episodes
with gzip.open(file_path, "rt", encoding="utf-8") as f:
    for line in f:
        episode = json.loads(line)
        print(f"Seed: {episode['seed']}, Steps: {len(episode['steps'])}, Rewards: {episode['rewards']}, Reason: {episode['reason']}")
        first_step = episode['steps'][0]
        print(f"  First step seat: {first_step['seat']}, action taken: {first_step['action']}, legal count: {len(first_step['legal'])}")
        break

License & Attribution

Licensed under the Apache 2.0 License.

@misc{tongits_benchmark_2026,
  title={Tongits Offline RL Benchmark: 10,000 Paired Imperfect-Information Trajectories},
  author={Rico Zuñiga},
  year={2026},
  howpublished={\url{https://huggingface.co/datasets/ricoz/tongits-benchmark}},
  publisher={Hugging Face}
}