We fine-tuned Qwen3.5-4B on Together AI to make decisions: give it context, a question, and 2–24 options, and it returns one answer letter. We're calling it tev1-4B-experimental.
This repo contains the data recipe, training example, and saved results so you can fine-tune your own decision model.
State Returns are allowed within 30 days. This purchase was 12 days ago.
Question Is this return within the allowed window?
Options A: Yes B: No C: Not enough information
Answer A
The latest recipe is called new v1 in the files. It starts from base
Qwen/Qwen3.5-4B and combines 37,840 unique training examples with
4,568 validation examples. The mixture covers language classification,
policy decisions, routing, and synthetic research classification. We use ordinary
LoRA supervised fine-tuning with Qwen's existing language-model output head.
- Checkpoint:
hassan/Qwen3.5-4B-v1-new-69617472-bdc3c2fc - Evaluated endpoint:
together/Tev1-4B-experimental - Saved results: 880/1,000 main decisions and 300/300 policy-transfer decisions.
These are reused development benchmarks, not untouched final tests. Endpoint
access depends on your Together account. See the run record
for evidence and limitations. The saved dataset has the counts above; uploaded-file
identity and the job's exact settings still need verification. The training example
uses the saved starting recipe: rank 8, one epoch, learning rate 5e-5, and a
2,048-token sequence limit.
Requires Python 3.12+ and uv. Run these commands from the repository root:
uv sync --locked
uv run python fetch_sources.py
uv run python build_all.pybuild_all.py runs the four builders in dependency order.
The older builder names remain because each is a dependency of the final dataset.
They build local intermediate data; new v1 is the dataset to train on.
Builders refuse existing output directories. See the dataset guide
for validation, provenance, and how to adapt the recipe.
Preview the training settings, then launch when ready:
uv run python examples/train_together.py
export TOGETHER_API_KEY='your-key'
uv run python examples/train_together.py --launch--launch uploads data/new-v1/instruction/train.jsonl and dev.jsonl and starts
a billed training job. Save the returned job ID and follow it in Together;
rerunning the command starts another job. For your own data, use --data path/to/instruction;
use --settings path/to/settings.json to override recipe settings.
See the training guide for the input format and deployment step.
Once you have deployed your fine-tune on Together:
export TOGETHER_MODEL='your-deployed-endpoint'
uv run python examples/decide.py examples/charge-dispute.jsonIf you copied .env.example to .env, use uv run --env-file .env to load it.
JEV_MODEL is also supported as an alias for TOGETHER_MODEL. More input examples
cover return policies, yes/no comprehension, and sentiment in examples/.
The example disables thinking and constrains the answer to the supplied option letters. It returns the selected letter, semantic key, and token logprobs. Logprobs are model preferences, not calibrated confidence.
Evaluate your endpoint on labeled records:
uv run python scripts/evaluate.py --provider together \
--model "$TOGETHER_MODEL" --records data/v2.1/records/test.jsonl \
--output outputs/my-evaluationUse a fresh holdout once evaluation results inform your training changes.
Code and original documentation are MIT licensed. Third-party datasets and model weights have their own terms; see source provenance. Training data and model weights are not bundled. This is an independent, Jev-inspired implementation and does not use Jev's answers as training labels.
Run uv run python -m unittest -v before contributing. See CONTRIBUTING.md.