Case study / Project archive
CodeSageZ
An earlier code-generation and fine-tuning experiment, distinct from the current CodeSage source workspace. Model quality is not established here.
Historical project; AI-assisted development. Individual and team contribution has not been independently verified.
The idea
The problem
Investigate whether a compact code model can be adapted for Python bug-fix completion while keeping data preparation, memory constraints, and evaluation explicit.
The implementation
What I built
The earlier experiment includes dataset preparation, memory-conscious adapter configuration, checkpoint recovery, and evaluation tooling. These are inspected implementation paths, not a newly verified model improvement.
- Python
- PyTorch
Features
- Prepare filtered Python bug-fix instruction/response examples from CommitPack.
- Configure a QLoRA adapter training path for Qwen2.5-Coder-1.5B-Instruct.
- Resume saved checkpoints and compare baseline and adapter evaluation outputs.
Architecture
- Python dataset preparation, Unsloth/Transformers fine-tuning, and separate CodeBLEU, HumanEval, and repository-context evaluation scripts.
- Training configuration uses short sequences, gradient accumulation, epoch checkpoints, and provenance helpers; the current source workspace is a separate portfolio entry.
Engineering decisions
- Keep training loss and text-similarity evaluation separate from evidence that generated repairs pass independent checks.
Validation and measured results
- Local review inspected training/finetune.py and training/MODEL_CARD.md. Historical GPU and score values were not reproduced or promoted into portfolio metrics.
Limits and unfinished work
- No new GPU training, held-out evaluation, contamination audit, or verified code-repair result was produced for this review.
- Older source can contain superseded descriptions. Numerical impact, adoption, awards, and deployment claims are not carried into this entry.
Evidence and sources
This project’s evidence remains local. Public source and downloadable receipts are not attached yet.