Changelog

All notable changes to ForgetLLM will be documented in this file.

The format is based on Keep a Changelog,
and this project adheres to Semantic Versioning.

[0.2.0] - 2025-06-30

Beautiful Training Log Viewer: Complete redesign with syntax-highlighted JSON display
- Color-coded JSON formatting (keys, strings, numbers, booleans)
- Dark theme with professional styling
- Metadata header showing training configuration at a glance
- Enhanced copy-to-clipboard functionality preserving original formatting
- Modal-based viewing without disruptive tab switching
Improved Training Sessions Layout: Professional card-based design
- Fixed model name extraction showing actual model names instead of "Unknown"
- Clean model name display (removed vendor prefixes like mlx-community/)
- Better spacing and responsive Bootstrap layout
- Enhanced visual hierarchy with icons and proper headers

Consolidated Update System: Reduced API calls from 4-5 per cycle to single call
- Single /api/dashboard/realtime endpoint every 10 seconds
- Eliminated redundant polling mechanisms
- Smart throttling preventing excessive updates
- Cache busting for JavaScript files
Enhanced Real-time Monitoring:
- Robust MLX process detection for accurate training status
- Optimized idle monitoring (10-second sleep when no training active)
- Improved data consistency across all API endpoints
- Better handling of optional fields in training metrics

Graceful Shutdown System: Complete process lifecycle management
- New ProcessTracker utility tracking all spawned processes
- SIGINT/SIGTERM signal handlers for clean termination
- 5-second graceful shutdown with force-kill fallback
- Automatic MLX process cleanup on exit
- Integration across all entry points (web server, training, CLI)
Enhanced Watchdog System: Extended timeout from 30 to 90 seconds for heavy LLM training
Improved Error Handling: Better fallbacks and error messages throughout the system

Unified Model Filtering: Consistent model lists across training and testing tabs
- Proper separation of base, CPT, and IFT models
- Published CPT model detection from HuggingFace cache
- Excluded unpublished training checkpoints from UI dropdowns
- 81 models (base + IFT) available in both training and testing tabs

Enhanced Training Sessions API: Better data extraction and formatting
- Fixed model name resolution from multiple data sources
- Improved checkpoint path parsing for numbered checkpoints
- Better handling of training metadata and configuration
- Comprehensive error handling and logging
Optimized Dashboard Data:
- Precise epoch calculation based on trained tokens vs dataset size
- Enhanced numeric formatting with proper precision control
- Better handling of small numbers (learning rates, etc.)
- Improved time calculations for elapsed and remaining time