Repository navigation
Train from a checkpoint verified against its NNUE - #12
Merged
arhancanli merged 1 commit intoOct 5, 2026
Merged
arhancanli merged 1 commit into
arhancanli merged 1 commit into
Conversation
arhancanli
marked this pull request as ready for review
October 5, 2026 18:37
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
The trainer can now start an isolated fine-tuning study from an existing PyTorch checkpoint instead of random weights. Both
--init-checkpointand--init-networkare required, and the checkpoint must reproduce the supplied NNUE byte for byte with the configured architecture. Invalid layouts, keys, dtypes, nonfinite weights and output aliases fail before candidate export. Source hashes are checked again after training, and.init.jsonrecords initialization and selected artifact hashes. Warm initialization defaults to LR 1e-5; the existing cold default remains 1e-3.Validation: all 62 local Python tests passed, including the original cold-run checkpoint regression. A real 512-hidden Metal zero-LR control visited all 69,613 frozen records, kept exactly the champion weights and NNUE, re-exported its checkpoint identically, matched 256 native integer evaluations and returned 16 legal searches. The reserved validation slot retained all 896 history-pruning games, configuration, seed and queue/settings hashes, and resumed the service.
This adds training mechanics for the later controlled study. The active controller remains on the explicit 512-hidden / 80M-fresh profile; no new learning result, strength gain or playing-engine promotion is claimed. The README reflects that installed profile.