Learning to Reason in 13 Parameters
Paper • 2602.04118 • Published • 6
Fine-tuned Llama 3.2 1B Instruct model for Texas Hold'em poker decisions with TinyLoRA. The adapter size is 470KB!
Training
- Base model: meta-llama/Llama-3.2-1B-Instruct
- Dataset: RZ412/PokerBench
- Method: TinyLoRA fine-tuning with unsloth-mlx
- LoRA config: r=256 (svd), u=1024, target_modules=[q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj]
- Training data: 50k preflop + 50k postflop samples
Performance
Evaluated on PokerBench test sets:
| Preflop | Postflop | |
|---|---|---|
| Base Model | 7% | 13% |
| Fine-tuned | 36% | 60% |
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for neopolita/TinyLoRA-TexasHoldEm-Llama-3.2-1B-Instruct
Base model
meta-llama/Llama-3.2-1B-Instruct