Fine-tune an open model. See if it got better.
Tuned Tensor is the tt CLI: one terminal for fine-tuning small open models on your GPU or on managed cloud. Every run compares the tuned model with its baseline on your examples and reports score changes and regressions, not just training loss.
$ npm install -g @tuned-tensor/cli
$ tt
tt v0.9.2
cloud · support-bot · no spec · model —
ctrl+c stop/clear · ctrl+d exit · /help commands · tab complete
Ask TT anything. Known commands run directly.
› init --model Qwen/Qwen3.5-2B
✓ Created tunedtensor.json — add your examples
› /mode local
✓ Workflow switched to local.
› doctor ▊Evaluation built into the run
Know whether your fine-tune beats the baseline.
Training loss measures fit to the training objective, not performance on your task. Tuned Tensor compares base and tuned outputs on representative evaluation cases. The local workflow creates a holdout when the spec has enough examples; for prebuilt data, it requires a test or validation split by default.
Baseline and tuned, side by side
Evaluate the same cases with the run baseline and fine-tuned candidate.
Metrics tied to the task
The local workflow measures pass rate, average score, exact match, or field-level JSON accuracy.
Regressions you can inspect
See which examples became worse and compare the expected, base, and tuned outputs.
An auditable run record
Keep the spec, selected evaluation cases, model artifact, events, and JSON report together.
Choose normalized exact match or field-level JSON scoring in the local workflow. Reports summarise evaluation on the cases you provide; they do not guarantee production performance.
A test-driven workflow
How a run works.
Your behaviour spec defines the task and the success criteria. The run report compares the tuned model with the baseline.
01
Define the task
Write the behaviour, examples, constraints, and base model in tunedtensor.json.
02
Evaluate the baseline
Run the baseline model against representative evaluation cases.
03
Fine-tune
Train a LoRA adapter with supervised fine-tuning.
04
Evaluate the tuned model
Use the same cases, generation settings, and scoring criteria.
05
Review the result
Inspect score deltas and failures before you keep or iterate on the model.
One CLI, two workflows
Your GPU or managed cloud.
Same spec, same commands, same paired report. What changes is who operates the GPU.
tt local
Run on hardware you control
The Apache-2.0 tt CLI trains and evaluates small models on Linux with a compatible NVIDIA GPU and sufficient VRAM, including NVIDIA DGX Spark. Run tt local ... without a Tuned Tensor account or hosted GPU queue.
- No Tuned Tensor account or hosted GPU queue
- Bundled Transformers/PEFT training, plus hooks for custom backends
- Specs, adapters, run events, and reports stored as local files
Base-model downloads use Hugging Face unless already cached. OpenRouter is optional for teacher labeling and LLM-judge scoring.
tt cloud
We operate the training infrastructure
The same CLI without the GPU. Plain tt commands (or tt cloud ...) run managed training when you do not want to run your own GPU. Start times depend on available managed GPU capacity.
- No local CUDA environment to maintain
- Managed training, teacher labeling, and model lineage
- Downloadable model artifacts and side-by-side run reports
Best fit
Tasks with clear success criteria.
Tuned Tensor is for focused application tasks with clear success criteria, not frontier-scale model research. Provide representative examples and score the outputs, and you can decide whether the fine-tune is worth using.
Structured extraction into strict JSON
Classification, safety, and triage
Routing and tool-selection decisions
Format, tone, and policy adherence
Measured on held-out data
Email triage: test pass rate rose 27.5 percentage points.
We fine-tuned Qwen 3.5 2B on 8,000 public training examples. On a 200-example held-out test sample, pass rate improved from 61.5% to 89.0% and average score improved from 0.537 to 0.862.
These results are specific to this task and evaluation setup. Test on data that reflects your production workload.
Base pass rate
61.5%
Tuned pass rate
89.0%
Pass-rate change
+27.5 pts
Held-out cases
200