Tuned Tensor Community Specs

Free, open-source Tuned Tensor spec files for small specialist language models.

This repository is intentionally simple. Each example is centered on one file:

tunedtensor.json

That JSON file is the behavior spec. It contains the task description, system prompt, guidelines, constraints, examples, and base model.

Optional notes can sit beside the spec when they are useful:

• dataset-card.md
• model-card.md
• eval.md

Template

Start from the simple template:

cp -R template specs/my-new-spec

Then edit specs/my-new-spec/tunedtensor.json. The optional cards can be deleted if you do not have a public dataset, model, or eval yet.

Use A Spec

Copy a spec locally:

cp specs/email-safety-triage-qwen2b/tunedtensor.json ./tunedtensor.json
tt eval -f tunedtensor.json

Then edit the JSON for your own task, examples, labels, or base model.

Validate Specs

Run the same check used by CI:

scripts/validate-specs.sh

Add A Spec

Add a folder under specs/ with:

tunedtensor.json
dataset-card.md
model-card.md
eval.md

Only tunedtensor.json is required. The notes are helpful when a public dataset, model, or evaluation exists.

Before opening a PR, read CONTRIBUTING.md. Do not commit model weights, private datasets, API keys, tokens, or .env files.

License

The repository content is licensed under the MIT License unless a file says otherwise. Linked datasets, base models, and fine-tuned models keep their own licenses.