Model-use summaries and limitations
These are educational model-use records for the implemented examples, not approvals for operational decisions. Full experimental configuration, splits, selected parameters, and measured outputs are saved with each run. Read data provenance before adapting.
| Family and modules | Intended learning use | Evaluation boundary | Principal limitation |
|---|---|---|---|
| Rules, M1 | Trace explicit decisions | Hand-checked boundary inputs | No model was trained |
| Forecasting, M5 | Availability, baselines, frozen evaluation | Chronological train/validation/test; rolling one-day observations | Historical bike demand is not current IT demand |
| Wine classifiers, M6/M8/M12 | Compare classical/neural approaches and checked inference | Group identical chemistry profiles across splits; validation-only selection | Classroom quality >=7 label is not certification; record groups are not verified physical identities |
| Wine clustering, M7 | Unsupervised selection and profile interpretation | Training scaling; validation silhouette; quality excluded | Cluster IDs are arbitrary, not quality ranks |
| SMS neural classifier, M9 | Training-only vocabulary, imbalance, error analysis | Deduplicated normalized real text; disjoint split IDs | Initial neural model missed all 30 positives despite 87.5% accuracy |
| Local LLM, M10/M12/M14 | Prompt/schema evaluation, checked interfaces, bounded tools | Fixed SmolLM2-135M-Instruct revision and bounded real examples | Initial M10 output schema success was 0/8; valid syntax would still not prove correctness |
| LoRA, M11 | Real adapter gradients, save/reload, and prompt masking | Eight tiny training steps; separate validation/final text IDs | Small completion-loss changes do not establish useful behavioral improvement |
| Adult classifier, M13 | Denominators, group audit, decision ethics | Provided benchmark boundaries with overlap removal; validation threshold | Historic income/category labels are limited and harmful if generalized to people; age remains an input |
Intended users and exclusions
Learners and instructors can inspect mechanics, change a bounded assumption, and explain measured results. These artifacts are not validated for grading individuals, employment, credit, medical use, incident response automation, quality certification, or unsupervised production actions.
The local Streamlit app has no production authentication, rate limiting, availability commitment, or durable operations process. Keep it local. Bounded agents use an outside-model allowlist, read-only tools, and a call budget; legal plans can still misunderstand intent. Never turn model text directly into a shell command.
Model identity and rights
The language-model identity, exact revision, license links, dataset citations, and reviewed archive hashes are in data sources. Remote custom code is disabled and safetensors are requested. Pinning identity does not establish safety or eliminate dependency risk.
Review before adaptation
Document the new population, decision, error costs, privacy requirements, available features, evaluation boundary, baselines, uncertainty, failure response, and accountable owner. Repeat relevant behavior and accessibility checks. Add a completed model card using the source template; do not copy this demonstration's performance into a new domain.