tomzv1/truco-cfr

Truco CFR: a Deep CFR bot for Argentine truco

A small neural network (MLP, 304 → 256 → 256 → 256 → 12, about 850 KB) that plays heads-up Argentine truco (to 30 points, no flor) with envido and truco. Trained from scratch by self-play with Deep CFR (Brown et al., 2019) on a single RTX 5070 Ti: no human games, no LLM, no fine-tuning.

Play against it in the browser: https://truco-cfr.netlify.app/ · Code: https://github.com/tomzvilla/truco-cfr

Files

File	What it is
model.onnx	Full network. Input features: uint8[batch, 304] (the encoding in engine/src/truco/encode.rs). Output logits: float32[batch, 12].
model.safetensors	Same weights for PyTorch (ScaledMlp in train/truco_train/model.py).
strategy.tcfrw	Same weights in the project's own format, read by the Rust engine (and the WASM build).
config.json	Sizes, action names and policy.

Policy: softmax of the logits over the legal actions only. Actions, in order: jugar0, jugar1, jugar2, envido, realEnvido, faltaEnvido, truco, retruco, valeCuatro, quiero, noQuiero, mazo (jugar{i} = play the i-th card of the hand sorted from strongest to weakest).

Parity: PyTorch, Rust, ONNX Runtime and WebAssembly outputs agree within 1e-4 (tests in the repo).

Results

Full matches to 30, duplicate format (each deal played twice with seats swapped):

Opponent	Matches	Bot wins	95% Wilson CI
Random legal moves	2000	90.9 %	[89.6; 92.1]
Hand-written heuristic	2000	63.4 %	[61.3; 65.5]
Claude Haiku (LLM reading the table as text)	100	59.0 %	[49.2; 68.1]

It learned to bluff: it calls envido with 7 points or less in 38 % of those hands, and 17 % of its truco calls come with a hand in the weakest 30 % (the heuristic: 0 %). Details in docs/eval-report.md.

Limitations

Heads-up only, no flor, house rules ("reglas de bar"). Evaluated against a heuristic and an LLM, not against strong human players. The Claude Haiku sample is small (100 matches).

License

Code and weights: MIT. Card art used by the web app is by Jose Angel Canabal and is not MIT.