Claude/loss weighting analysis prt3o #36
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Training improvements: - Add cosine annealing LR schedule with warmup (first 10% of epochs) - LR decays from initial rate to 0.1× over training - Display current LR after each epoch for monitoring Training results (3 iterations, 300 games total): - Dataset grew from 1,453 to 16,341 samples (11× increase) - Movement policy loss: 4.45 → 2.76 → ~2.5 (44% improvement) - Drafting policy loss: 3.99 → 3.51 → ~3.3 (17% improvement) - Value loss consistently near zero in later epochs Model trained on 16,341 samples: - 2,640 drafting positions - 13,701 movement positions - Final epoch losses (Movement): P=2.76, V=0.001, O=0.43, S=0.01 The learning rate schedule enables better convergence by warming up gradually then decaying, preventing overshooting early and enabling fine-tuning later in training.
- Add --num-filters and --num-blocks command-line arguments - Pass architecture parameters through to train.py - Enables training with different model capacities - Defaults: 64 filters, 4 blocks (original architecture)
Model architecture increased to 2× capacity: - Filters: 64 → 128 (2× wider) - Residual blocks: 4 → 6 (50% deeper) - Total parameters: ~4× larger Bootstrap training results (20 epochs, 9669 samples): - Movement policy: 4.37 → 2.27 (-48% improvement) - Drafting policy: 4.12 → 3.17 (-23% improvement) - Value loss: → 0.0002 (near-perfect convergence) Comparison to smaller model (64 filters, 4 blocks): - Movement policy reached 2.27 vs 2.76 (21% better) - Much faster learning and lower final losses - Demonstrates value of increased model capacity Training continues with 3 iterations of self-play... Note: Training still in progress, models will improve further.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
No description provided.