togethercomputer / tev1

# tev1-4B-experimental
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.
- Learn how to train your own classifier for $17
- Full model weights
 [code] 
## The 4B run
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.
## Fine-tune it
Requires Python 3.12+ and uv.
Run these commands from the repository root:
 [code] 
`build_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:
 [code] 
`--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.
## Try your model
Once you have deployed your fine-tune on Together:
 [code] 
If 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:
 [code] 
Use a fresh holdout once evaluation results inform your training changes.
## License and contributions
Code and