Training improvements: LoRA tuning, early stopping, loss normalization#23
Merged
Training improvements: LoRA tuning, early stopping, loss normalization#23
Conversation
- LoRA: rank 64→16, alpha 128→32, dropout 0→0.05 (research-backed) - Epochs: 3→2 with early stopping (patience=5, threshold=0.01) - Enable mid-epoch validation every 250 steps (was disabled) - Add _check_early_stopping pure helper with proper edge case handling - Check early stopping at both mid-epoch and epoch-end validation - Log early stopping events to train.log and metrics.jsonl
- _iter_training_batches returns (batch, total_tokens, completion_tokens) - Training loss divided by completion token count for interpretable values - Validation loss accumulated as total_loss/total_completion_tokens - metrics.jsonl logs both train_loss (per-token) and train_loss_total (raw) - final_loss and logger.info use per-token values
- Parses metrics.jsonl (train, val, early_stop entry types) - Two-subplot layout: loss curves + throughput - EMA smoothing for noisy train loss (configurable alpha) - Marks epoch boundaries, best val loss, early stopping - Handles both old (sum-reduced) and new (per-token) formats - CLI: --metrics, --output, --dpi, --ema-alpha - 12 parser tests
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 76afa9afec
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
- Fix mypy type error in step_loss_per_token assignment - Guard best_val_loss lookup against missing exact match (P1) - Persist early stopping state across checkpoint resumes (P2)
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
1 participant
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.
Summary
Fixes severe overfitting observed in Qwen3-8B training (val loss diverged from ~94 to ~130 by epoch 3). Changes based on deep research into Qwen3 LoRA best practices.
eval_interval_steps=250(was 0/disabled!)scripts/visualize_tinker_training.pyparsesmetrics.jsonland plots loss curves with epoch boundaries, best val loss, and early stopping markersFiles Changed
configs/training.yamlsrc/training/train_tinker.py_check_early_stoppinghelperscripts/visualize_tinker_training.pytests/test_train_tinker.pytests/test_visualize_tinker_training.pyTest Plan
ruff check)