STC_S1Mini_Model

First start: open /admin/ → Updates → Download; see the web UI guide. Startup and requests load downloaded files only; an empty installation keeps health/admin available and returns 503 for inference. Import the model with the Ollama guide before downloading its tokenizer.

The one place S1-mini is served. S1-mini is a 0.6B text normalizer for speech-to-text output (Qwen3 fine-tune by Superwhisper). Projects send it a raw transcript over HTTP and get clean text back: fillers removed, self-corrections resolved, punctuation and capitals added, spoken numbers, times and emails written out. It is not a chat model. English only. It has the same admin page, tokens, rate limits, metrics, settings and update checks as the other model services.

S1-mini always runs through Ollama. Local URL defaults to http://127.0.0.1:11434; Compose defaults to http://host.docker.internal:11434. Explicit empty URL is an error. Start Ollama and prepare s1-mini:f16 first; see pinned setup. The model-card system prompt stays unchanged, with greedy decoding and thinking off. Only the admin Updates Download action downloads the pinned tokenizer. Ollama owns model memory; API replicas keep independent queues and answer caches.

API

Call	Purpose
POST /v1/normalize	{"text", "styling", "structure", "context", "stream"} → {"content", "finish_reason", "model", "usage", "duration_ms"}.
GET /health	{"status": "ok", "model", "ready", "num_thread"}; ready is true once the model is loaded. No token needed.

styling: casual, semi-casual, semi-formal (default), formal. structure: prose (default) or lists. context: general (default) or email. Other values are refused with 422: the model garbles output for values it was not trained on. Any other field (messages, tools, temperature, ...) is refused with 422 too.

Example: so um i need to like send the the report by uh friday no wait make that thursday becomes So I need to send the report by Thursday.

An empty content is a valid result (a transcript that holds only fillers).

Long text. Input is cut at sentence ends into passes of at most Input tokens per pass (SERVICE_CHUNK_TOKENS, default 800; the model card advises under about 1000) and each pass is normalized on its own. One sentence longer than the limit is sent whole. Output joins with a space (a newline for lists).

Streaming, per sentence. With "stream": true the reply is application/x-ndjson. One line per finished sentence, not per token: {"type": "sentence", "text": "So I need to send the report by Thursday. "}, then {"type": "done", ...} holding the same body a call without stream returns. Each text keeps the whitespace that follows it, so joining the pieces gives the answer. The first sentence arrives when the model has finished it. A refusal before the first sentence (a full queue) is a normal HTTP error; a failure after it is sent as {"type": "error", "status", ...} and ends the stream. A kept answer (see below) is streamed back sentence by sentence.

Cache. Answers kept per replica (SERVICE_CACHE_SIZE, default 256, 0 = off) keeps the last answers for identical requests and sends them again without generating (X-Cache: hit). Decoding is always greedy, so every request is kept. New weights empty the cache. The cache is per replica, so with N replicas the hit rate is about 1/N. stream does not change the key.

Tokens. Project scopes are removed: token create/edit accepts only omitted or null projects (for shared-editor compatibility); any list returns 422. Inherited scopes are ignored and hidden.

Admin page

/admin/ (password SERVICE_ADMIN_PASSWORD): tokens with rate limit and expiry, usage numbers, CPU threads, the Settings card (update check hours, in-flight cap, answers kept, input tokens per pass), and the Updates card (Check now shows the local import digest with no online version; Download fetches the pinned tokenizer). The shared data-volume marker invalidates every replica's loaded handle and answer cache. CPU threads are sent to Ollama as num_thread with every request; unset lets Ollama decide. Set it to the physical core count. /metrics (admin password) is Prometheus text per token. s1mini_request_seconds is a histogram of how long each /v1 call took on the replica that answers (to the first byte for a streamed reply); it starts from 0 when the replica restarts.

Run

docker compose --env-file .env.compose up -d --build        # http://<host>:29280

Without Docker: run.bat (uv, listens on 127.0.0.1:29280). Ollama must already run. If Ollama is unavailable or its model is missing, health/admin still answer with readiness false; normalization returns 503 with a admin Updates Download hint. The next request retries automatically. See deployment.

Tests

uv sync
uv run pytest

Limits worth knowing

• Up to SERVICE_OLLAMA_PARALLEL generations at once per replica (default 2); extra calls wait, and more than the in-flight cap get 429.
• The rate limit is per replica.
• A pass has a 120 s timeout. Backend failures return 503 before streaming or an error event after streaming starts.
• A style that drops end punctuation (casual) can leave no full stop between two passes of a long text.
• Licence: Apache 2.0 plus a naming clause. Wherever the model is named it must be "S1-mini" by "Superwhisper". Read the LICENSE on the Hub before redistributing.

Try it from the admin page

Open /admin/ and use Try it to call the public API with a service token held only in memory. Runs use normal token scopes, limits and caches; the panel shows HTTP status, latency, responses, cancellation and a curl equivalent with a token placeholder. See the web UI guide.

Normalize sends transcript text plus styling, structure and context to POST /v1/normalize. Streaming appends sentence events as they arrive. This cleans text with S1-mini by Superwhisper; it does not transcribe uploaded audio or act as a general chat model.

Documentation

• Purpose: input, output and what it is for
• Ollama setup
• Web UI
• Deployment
• Changelog
• Docker on another machine
• Docker on this PC
• Run without Docker
• Test it from the command line

Admin pages and navigation

Open /admin/ and sign in. Overview opens first; each sidebar destination has its own URL fragment and supports browser Back/Forward and direct links:

• #overview: health, model readiness, replicas, active tokens and request counts.
• #playground: service-specific request forms and streamed or media responses.
• #tokens: create, edit, expire and revoke scoped access tokens.
• #runtime: CPU threads where supported, replica information and live/restart settings.
• #updates: check versions and explicitly download weights.
• #connections: caller addresses and public endpoints.
• #models and #voices: project models and cloned voices, shown only when supported.

Collapse/Expand switches between labeled sidebar and icon rail on desktop and mobile. The non-secret sidebar preference is saved per service in this browser. Navigation preserves form drafts and active Playground requests; sign-out still clears credentials and cancels work. Page names appear once in the main header. No preview data or cross-service selector ships in the real admin UI. Existing authentication, quotas, settings precedence and explicit-download policy remain in effect.