ml-research-lab

# ML Research Lab Use this skill as the compact router for ML work. It is derived from an audit of `synthetic-sciences/openscience` at commit `531467c`, but does not require running OpenScience or loading its full 250+ skill set. ## Operating Loop 1. Freeze the question as a measurable hypothesis. 2. Identify dataset provenance, labels, splits, leakage risks, and regeneration cost. 3. Pick the smallest baseline that can disprove the idea. 4. Define metrics before training. For release claims, require train/val/test split, no test-set model selection, and multi-seed proof when cost permits. 5. Run or wire experiment tracking before long jobs start. 6. Save artifacts: config, command, data manifest, metrics JSON/CSV, logs, model hash, and a short conclusion. 7. Compare against baseline, then keep/discard the change from evidence. ## Domain Routing - Dataset or scrape cleanup: start from data quality, deduplication, leakage checks, train/eval splits, and regeneration notes. - Classical classifier or tabular baseline: use scikit-learn-style pipelines with preprocessing inside the pipeline and stratified splits for classification. - Model debugging or trust: add SHAP/explainability for feature importance, leakage, bias/proxy features, and misclassified samples. - LLM fine-tuning: prefer JSONL chat format, data validation, LoRA/QLoRA baseline, and tracked runs before scaling. - Single-GPU fast LoRA/QLoRA: consider Unsloth only after checking hardware, CUDA, mod...