# Easy Jev fine-tune

This repository accompanies the article **“I Built My Own ‘Jev’ and Fine-Tuned
It on Banking Data.”** It contains the complete Banking77 classifier, the LoRA
adapter used for the published results, tests, and machine-readable evidence.

The Pages site also preserves the earlier **small-LLM LoRA experiment** as a
separate article at `small-llm.html`. Its code, data, notebook, and evidence live
under `experiments/small-llm/` so the two experiments remain clearly separated.

This is a small, local, Jev-shaped experiment—not TypeSafe's Jev and not a
reimplementation of its architecture. It explores the same useful product idea:
when a product needs a bounded decision, return typed probabilities instead of
generating prose.

## Try the trained model

Python 3.12 is the tested version.

```bash
git clone https://github.com/shmcsensei/easy-jev-fine-tune.git
cd easy-jev-fine-tune
python3.12 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -r requirements.txt

python banking77.py predict "Why was my card payment declined?"
```

The committed adapter means prediction does not require retraining. The first
run downloads a pinned revision of `HuggingFaceTB/SmolLM2-135M-Instruct` from
Hugging Face.

## Reproduce the training

The published adapter used all 10,003 official Banking77 training examples for
one epoch. The official 3,080-example test split remained untouched until final
evaluation.

```bash
python banking77.py train \
  --examples-per-class 999999 \
  --epochs 1 \
  --output adapter-banking77-reproduced

python banking77.py evaluate \
  --adapter adapter-banking77-reproduced \
  --threshold 0.70 \
  --output evidence/reproduced-evaluation.json

python banking77.py benchmark \
  --adapter adapter-banking77-reproduced \
  --samples 1000 \
  --warmup 30 \
  --output evidence/reproduced-latency.json
```

For a faster smoke test, omit `--examples-per-class`; the development default
uses a balanced 20 examples per intent for three epochs.

Hardware, package versions, and nondeterministic operations can cause small
differences. Latency will vary substantially by device.

The reported 0.70 threshold and threshold curve are descriptive analysis of the
test set, not a production-validated operating point. A real deployment must
choose its threshold on separate validation data and reserve a final test set.

## Verify the lightweight tests

```bash
python -m unittest -v
```

## Repository map

- `index.html` — the published article
- `small-llm.html` — the earlier small-LLM fine-tuning article
- `BLOG_POST.md` — the article in Markdown
- `banking77.py` — training, prediction, evaluation, and benchmarking
- `finetune.py` — shared model/device helpers
- `adapter-banking77/` — the trained LoRA adapter used in the article
- `evidence/` — before/after evaluation and latency JSON
- `test_banking77.py` — policy-routing and benchmark-helper tests
- `experiments/small-llm/` — code and evidence for the second article

## GitHub Pages

The workflow in `.github/workflows/pages.yml` publishes the site on every push
to `main`.

## License and third-party material

The repository's original code, prose, and site assets are released under the
[Unlicense](LICENSE), an OSI-approved public-domain dedication with a permissive
fallback. Banking77 is CC BY 4.0, and SmolLM2 is Apache-2.0. The committed
adapter and Banking77-derived evidence retain Banking77 attribution. See
[THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md) for details.

This is a research demonstration. Do not use it for autonomous financial,
fraud, identity, eligibility, or account-access decisions.