LLM-course/TRM_d48_L1_H2_C4_35k_LegalW0p5_ELO_finetuned

Chess model submitted to the LLM Course Chess Challenge.

Submission Info

• Submitted by: janisaiad
• Parameters: 43,104
• Organization: LLM-course

Model Details

• Architecture: Tiny Recursive Model (TRM) - looping recurrent transformer (cycle-shared weights)
• Vocab size: 148
• Embedding dim: 48
• Layers: 1
• Heads: 2
• Cycles: 4

TRM note: this is a looping TRM model — at inference/training time we run the same transformer stack for 4 recurrent refinement cycle(s) (weights are shared across cycles), which increases compute/reasoning depth without increasing parameter count.

ELO Fine-Tuning

This model has been fine-tuned on Lichess games filtered by ELO rating (1200-1400) to improve chess playing strength and win rate. The fine-tuning focuses on learning from games played at intermediate skill levels, optimizing the model to make stronger moves and win more games.

Fine-tuning details:

• Dataset: Lichess games filtered by ELO (1200-1400 range)
• Objective: Maximize win rate and chess playing strength
• Training approach: Supervised learning on high-quality game sequences