FAQ

Does Retoque run my model?

No. You download the base model from Hugging Face and run it on your own hardware or a host you control. Retoque distributes the adapter, the evidence and the license, and the CLI loads everything locally. There is no hosted inference.

Is the registry key an OpenAI, Anthropic or Hugging Face key?

No. A key starting with rtq_ authenticates you to the Retoque registry for license checks and downloads. It can't call any model and is never sent to a model host.

Which models are supported?

One base model at first release: Qwen/Qwen2.5-1.5B-Instruct at commit 989aa7980e4cf806f80c7fef2b1adb7bc71aa306. The live list is on Supported models and at GET /api/v1/models. Other sizes, revisions and quantized variants are not supported.

Can I use a patch on a different revision or a quantized model?

Not as a supported install. The CLI refuses anything except the exact repository and commit unless you pass --experimental-override, which is recorded in retoque.lock. Results for that combination are unknown.

Does installing a patch change my base model?

No. The base weights are only read. A patch is a separate adapter file, and retoque disable returns you to the pure base model without reinstalling anything.

How do I know a patch works?

Read its evaluation on the patch page: the held-out baseline and patched counts out of the full suite, the prompt-only control, the regressions introduced, and whether rollback reproduced the base outputs. The evidence state says how far it has been checked. See Evaluation methodology.

Why can't I see the held-out cases?

Publishing them would let anyone tune a later version against the benchmark. The suite's sha256 is published instead, and reviewers with access can reproduce the run. Cases shown on a patch page come from the validation split.

What does the signature prove?

That a trusted publisher signed these exact bytes. It says nothing about output quality. See Artifacts and verification.

Is the onchain license live?

The registry and license market contracts are written and tested, and run on a local development chain. They are not deployed to Robinhood Chain mainnet. Mainnet waits on an independent security review. Free click-through licenses work today. See Robinhood Chain.

Can a license be revoked after I download?

Revoking or suspending a license stops new downloads. It can't recall files already on your disk. The license terms still govern how you use them.

Can I resell a license?

Onchain licenses are bound to the buyer's wallet inside the contract and are not tokens, so there is nothing to transfer. Free licenses are tied to your account.

Can I serve a patch to my users?

Only if the version's terms allow it. Check the commercial scope on the patch page. Technically, serve it with static adapters as described in Serving with vLLM, which is untested for published patches.

Can I build my own patch?

Yes, in the creator studio. Upload a JSONL dataset with families, train with the fixed configuration, evaluate, and submit for review. See the Creator guide. Training runs on operator hardware for now. Hosted GPU workers are planned.

Can I tune the training hyperparameters?

No. Every studio run uses one bounded LoRA configuration, so results across patches are comparable and reviewable.

Is there a way to remove an installed patch?

There is no uninstall command in CLI 0.1.0. retoque disable stops it from loading. To free the space, delete ~/.retoque/patches/<slug>/<version> yourself.

Where can my assistant read these docs?

Every page is available as Markdown at /docs/md/<page>. /llms.txt indexes them, and /llms-full.txt has all of them in one file. GET /docs/search?q=<query> returns ranked sections as JSON.