[r33drichards]

Summary

This PR adds comprehensive documentation for fine-tuning the Laya vision model on custom datasets, including both local training and Modal-based GPU training approaches.

Key Changes

• New guide document (site-docs/how-to/fine-tune.md): Complete fine-tuning workflow covering:

  • Data format specification (JSONL with images directory structure)
  • Detailed field descriptions and validation rules
  • Best practices for dataset preparation
  • Local training with Python using freeze options ("head", "last_n", "full")
  • Modal-based training on A100 GPUs with per-epoch evaluation
  • Temperature fitting for calibration
  • Model saving and sharing to Hugging Face Hub
  • Usage examples for the fine-tuned model

• Updated navigation (mkdocs.yml): Added "Fine-tune on your own dataset" link to the How-to section

• Updated overview (site-docs/how-to/overview.md): Added reference to the new fine-tuning guide with brief description

Notable Details

• Provides clear comparison table of freeze options with use cases and performance characteristics
• Includes practical code examples for both local and Modal training workflows
• Documents dataset preparation best practices (question phrasing, split management, validation set sizing)
• Explains temperature fitting for maintaining meaningful confidence scores after training
• References related concepts and implementation details in laya/vlm_train.py

https://claude.ai/code/session_011MHDFrSKQo5r47X9SuTrpT