Module 11 · Neural and language models
Teach and adapt Module 11
Link the entry page as additional reading/practice. Its reading-site route is /lessons/11-fine-tuning/ relative to the served site base. For a repository link, select a tested commit or release.
Facilitation and access
Use the same outcomes and written evidence for in-person/asynchronous learners. Begin with the foundation bridge, predict-before-run questions, and the shared script/notebook. Discussion has a solo equivalent. Optional depth is skippable, and GenAI is not required.
Interpretation can use named values and exact tables; code/artefacts demonstrate execution or authoring when those are assessed. A screenshot cannot substitute for an explanation, and a script does not prove notebook-interface operation. Verify the learner's actual assistive-technology route and provide private help.
Adaptation decisions
A real organization needs consent/rights, a data-retention policy, versioned adapters, regression evaluation, and rollback. This tiny adaptation is not a production support model.
When changing a dataset, define its license/provenance, observation unit, target if any, availability, split, and limits first. Update preprocessing, inference contracts, examples, feedback, and tests together. Preserve the original source and disclose transformations. Do not label a tiny mechanics check as a credible production evaluation.
Outcome/evidence alignment
- M11.1: Distinguish prompting, inference, supervised fine-tuning, and LoRA.
- M11.2: Prepare completion labels with prompt-token loss masking.
- M11.3: Update a small adapter on training-only real annotations.
- M11.4: Compare predeclared base/adapted held-out completion losses without promising improvement.
- M11.5: Save/reload adapter arrays with the exact model revision and disclose limits.
Course alignment opportunities: CLO4, CLO5, CLO8. Advisory computing-outcome opportunities from the planning register: A1, A2, A3, A4. These are opportunities for evidence, not proof of master's-level attainment or accreditation compliance. The original-module crosswalk remains in the previously prepared planning resources. Confirm adopted program criteria before formal use.
Verification limits
Actual framework training/inference, the public script, and notebook are exercised in the Linux CPU job. Windows/macOS framework execution is not inferred from that run. Exact counts/results are in the verified results table and commit-specific CI. Consult the accessibility evidence log for human interface-review coverage. Public feedback is for self-study; local grading/AI-use rules remain the adopting instructor's responsibility.
Continue the learning route
Use the course review and pilot protocol. Review a changed-case explanation, not only passing visible checks. Preparation and activity times need observation with actual learners.