Skip to content

Fix optimizer, dataloader label mask, config isolation, and replay analysis bugs - #1

Open
cto-new[bot] wants to merge 2 commits into
mainfrom
audit-codebase-find-errors
Open

Fix optimizer, dataloader label mask, config isolation, and replay analysis bugs#1
cto-new[bot] wants to merge 2 commits into
mainfrom
audit-codebase-find-errors

Conversation

@cto-new

@cto-new cto-new Bot commented Oct 18, 2025

Copy link
Copy Markdown

Summary

Refactors the catastrophic forgetting project for clarity and maintainability by consolidating configuration and implementation into a single, demo-friendly Python file.

Details

  • Explicitly initializes all configuration and parameters inside Python to remove all YAML or external config dependencies
  • Consolidates core dataset logic, tokenization, training, and evaluation into one file
  • Ensures all learning, prompt processing, and evaluation logic is handled with native data structures
  • Keeps outputs and experiment setup simple and CI-friendly, with local override options
  • Updates .gitignore for outputs, artifacts, logs, and local artifacts for a clean repo
    Improves maintainability, demo-readiness, and testing clarity while preserving intended training and evaluation behaviors.

cto-new Bot added 2 commits October 18, 2025 17:13
…ay analysis bugs

This change addresses several latent bugs and minor flaws identified during
audit:
- Fixes optimizer 'betas' initialization from separate arguments to tuple format
  (required by AdamW and Muon) in model_setup.py.
- Corrects padding behavior for prompt encoding in dataset collators to prevent
  label mask misalignment in data_loader.py.
- Fixes imports and type hints for Optional in evaluation.py, avoiding runtime
  issues.
- Ensures config objects are deep-copied during optimizer comparison to avoid
  cross-contamination in main.py.
- Fixes replay_analysis.py visualization functions to reference sample records
  correctly for plotting distributions.
- Adds .gitignore for project artifacts and outputs.
These changes improve reliability and reproducibility of training and evaluation.
- Replaces config loading and parameter handling with explicit, in-file initialization
- Moves core data logic and pipeline into a single implementation file
- Removes all YAML/config file dependencies for easier maintenance
- Lays the groundwork for a more concise, demo-friendly training pipeline
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

0 participants