Tuned Tensor

Tuned Tensor Documentation

Tuned Tensor turns a behaviour spec into a fine-tuned small open-weight model and a paired report showing whether it beat the run baseline on representative evaluation cases.

The tt CLI is the interface: a laptop-local agent, CUDA fine-tuning on hardware you control, and OpenAI-compatible serving. There is no hosted training workflow. Specs live in tunedtensor.json.

Key Concepts

Behaviour Specs

A behaviour spec is a structured description of what you want your model to do. It lives in your project as tunedtensor.json. It includes:

  • System prompt — the persona and role of the model
  • Guidelines — rules the model should follow
  • Constraints — things the model must not do
  • Examples — input/output pairs demonstrating desired behaviour
  • Base model — which open-weight model to fine-tune

Runs

A run snapshots the behaviour spec, prepares training and evaluation data, fine-tunes a LoRA adapter, and compares the baseline with the tuned candidate on matched cases. Artifacts stay on the execution host.

Paired Evaluation

Tuned Tensor evaluates the baseline and tuned candidate on the same cases and reports score and pass-rate deltas plus per-example regressions. The local workflow creates a holdout when the spec has enough examples; prebuilt data prefers validation and then test by default. Results are evidence on those cases, not a guarantee of production improvement.

Local Serving

Fine-tuned artifacts can be served with tt models serve or tt serve active after a passing generalRegression gate. See Local Training.

Workflow

  1. Define — Create a behaviour spec describing what your model should do
  2. Run — Evaluate the baseline, fine-tune the candidate, and evaluate it on matched cases
  3. Inspect — Review the paired report, aggregate deltas, and per-example regressions
  4. Serve — Use tt models serve for a verified adapter
  5. Iterate — Refine your spec and run again

Certified base models

Local training currently certifies text SFT with LoRA/PEFT on CUDA:

  • Qwen/Qwen3.5-2B — default for most application tasks
  • nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 — Spark-class unified memory
  • meta-models/Muse-Glimmer-30B — vision-language text tower (vision frozen)

See Model recommendations for when to pick each one.

CLI

Install with curl -fsSL https://tunedtensor.com/install.sh | sh, or npm install -g --ignore-scripts @tuned-tensor/cli. Open source (Apache-2.0) on GitHub. See the TT CLI docs and the Quickstart.