togethercomputer/tev1 on GitHub
- Stars
- 23
- Forks
- 4
- Language
- Python
- License
- MIT
- Last push
- 24 Sep 2026
Post on X
- Views
- 59,086
- Likes
- 690
- Reposts
- 42
Read from GitHub and X on . Counts change daily.
Reported by the author
- Training cost
- $17
Tev1 cost $17 to train!
x.com - Serving price
- $0.042/M in
We're making it available on Together serverless at $0.042/M input & $0/M output.
x.com - Dev eval score
- 880/1,000
Saved results: 880/1,000 main decisions and 300/300 policy-transfer decisions.
github.com
Tev1-4B-experimental is Together AI’s Jev-inspired decision model on an open base: give it context, a question, and 2 to 24 options, and it returns one answer letter rather than free text. The repo starts from base Qwen3.5-4B and fine-tunes it with ordinary LoRA supervised fine-tuning on top of Qwen’s existing language-model head. The training mixture, called “new v1” in the files, has 37,840 unique training examples plus 4,568 validation examples covering language classification, policy decisions, routing, and synthetic research classification. The README says it does not use Jev’s answers as training labels.
Together reports the training run cost $17, and it serves the model on Together serverless at $0.042 per million input tokens with free output tokens. The repo’s saved results show 880 out of 1,000 correct on its main decision set and 300 out of 300 on a policy-transfer set. Together notes these are reused development benchmarks rather than untouched final tests, and that the uploaded files and the job’s exact settings still need verification. The training example uses the saved starting recipe: rank 8, one epoch, a learning rate of 5e-5, and a 2,048-token sequence limit. A tutorial by Hassan on X walks through fine-tuning your own version.
