Open Model Lab
SFT Pipeline + Data Quality
Measure the behavioral difference between a base model and an instruction-tuned model.
Gate status
- Target
- November 2026
- Status
- scheduled
- Report
- Results pending
Success criterion
The work does not stop at fine-tuning; it shows measurable behavior changes, including regressions.
Execution summary
Dataset preparation completed on September 24, 2026: 200 training examples, 40 validation examples, and a 24-example smoke subset. Training resumes in November 2026 with a real LoRA update, adapter save, and reload on M4 Max.
Focus
- Build an SFT pipeline.
- Define and clean instruction data.
- Compare base vs SFT behavior using the same eval harness.
- Measure eval regressions: what improved, what got worse?
- Create dataset and model cards.
Completed preparation
- Pinned Qwen3-8B MLX artifact and a recorded compatibility check.
- Validated SFT schema, completion-only loss masks, and tokenizer checks.
- Versioned pilot data, split manifest, dataset card, and preparation report.
Expected outputs
- training/sft module.
- Base/SFT behavior comparison table.
- Dataset card for the versioned SFT pilot.
- Report: what behavior does SFT improve and what can it damage?
Stage completion decision
Is the SFT checkpoint reliable enough to serve as the reference for preference optimization?