Datasets:
pretty_name: Range Reader v0.1 — CFR Self-Play Hand Histories
license: apache-2.0
annotations_creators:
- machine-generated
source_datasets:
- original
size_categories:
- 100K<n<1M
tags:
- poker
- GTO
- Game
- texas-holdem
- hand-history
- open-hand-history
- self-play
configs:
- config_name: default
data_files:
- split: train
path: out_min-2_max-6.jsonl
task_categories:
- reinforcement-learning
range-reader-v0.1
613,399 No-Limit Hold'em hands in Open Hand History format (spec 1.4.7), one JSON record per line with a blank seperator line. Hole cards are shown for every seat. This is the training set for range-reader, which predicts a villain's hole cards from the action.
Generation
Self-play from rs-poker's arena:
rsp arena generate ./examples/configs -o out_min-2_max-6.ohh -n <hands>
Cash games, 5/10 blinds, randomized stacks, 2- to 6-handed (even split). Five agents, four built on counterfactual regret:
| Agent | Type | Notes |
|---|---|---|
CFR-Configurable |
real-time CFR | known-card estimator, coarse bet sizes |
GTO-Experiment |
real-time CFR | — |
6Max-RFI-GTO, Pekarstas-6max-RFI |
preflop CFR chart | fixed open ranges |
RandomPotControl |
pot-control heuristic | not CFR; call odds scale with pot, plus noise |
Caveats
The CFR is a real solve, but bounded — strong on the immediate decision, thin past it:
- Up to 2048 ms per decision, 128 root iterations, regret-convergence early stop, root width set to the core count.
- Lookahead thins fast — one iteration and width 1 below the root, so multi-street planning is shallow.
- Coarse action abstraction: a handful of bet sizes (~2.5x raises, ⅓/⅔/full pot, shove).
- Charted agents play fixed preflop ranges; the pot-control agent is a heuristic.
Expect competent, varied poker — short of a full-tree Nash equilibrium.
Schema
Each line is {"ohh": {...}}: table and blinds, players with seats and stacks, rounds of streets, and per-street actions (post, deal, fold, call, raise, all-in). Dealt cards carry the cards field. See the OHH spec.
License
Apache-2.0.