Project brief
From scenario and baseline to reproducible evidence.
Post-train an open coding model with traceable, permitted data and demonstrate improvement on a held-out programming task without evaluation leakage.
The handbook starts with a runnable minimum, then adds real constraints. The final presentation should explain the method, results, failures, limitations, and next steps.
Evaluation focus
Use metrics to make trade-offs visible.
Project handbook
Project brief
Choose a narrow, automatically verifiable task such as Python bug fixing, unit-test generation, SQL, API use, code review, or repository question answering. Use a precisely versioned open model and compare it with the untouched baseline.
Minimum version
Create train, validation, and never-seen test splits. Start with a 7B-class model and LoRA, QLoRA, or an equivalent parameter-efficient method.
- Record every data source, license, cleaning rule, and processing date.
- Run deduplication and leakage checks before training.
- Evaluate with tests, compilation, linting, execution, or a clear reference standard.
- Compare target-task gains with retained general coding ability and resource cost.
Safety and release
Run generated code in an isolated test environment and never expose production keys, private repositories, or internal services. State clearly whether the model, data, and adapters may be redistributed under their original licenses.
Resource note: External code, models, datasets, and platforms remain subject to their own licenses, terms of service, and applicable law.