penva/affine-l

Affine-S3-GAME-Improved

Fine-tuned version of WebScraper991923/Affine-S3 with improved GAME (OpenSpiel) performance for Bittensor Subnet 120 (Affine).

Model Details

• Base Model: WebScraper991923/Affine-S3 (Qwen3-4B)
• Training: LoRA fine-tuning on 7,071 MCTS-generated game examples
• Target: Improved strategic game-playing for Affine evaluation

Training Details

• Method: LoRA (r=32, alpha=32)
• Data: 7,071 examples from MCTS self-play across 9 games:
  • checkers (2,702 examples)
  • gin_rummy (1,896 examples)
  • othello (1,209 examples)
  • quoridor, phantom_ttt, hex, dots_and_boxes, leduc_poker, liars_dice
• Epochs: 2
• Final Loss: 0.024

Performance

Benchmark	Base Model	This Model
GAME Accuracy	~30%	76%
LGC	99.9%	99.9% (preserved)

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("altro/Affine-S3-GAME", torch_dtype="bfloat16", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("altro/Affine-S3-GAME")

Affine Competition

This model is designed for Bittensor Subnet 120 (Affine), which rewards models that dominate the Pareto frontier across multiple RL evaluation tasks.