centinela-sentiment

centinela-sentiment preview

finanzas text-classification classifier gate: passed

Hugging Face: astromesh/Centinela-Qwen3-4B

Preliminary

This model is defined in the catalog but not yet published. The revision SHA and eval metrics below are placeholders until the first release.

Contract

• Labels: positivo · neutral · negativo
• Validation: constrain_label

Aliases

Alias	Revision
prod	v0.1

Revisions

Version	Base model	Dataset	Gate	macro_f1	invalid_rate
v0.1	Qwen/Qwen3-4B	centinela-sentiment-es@1	passed	0.00	0.00

Revision detail

v0.1

• Base model: Qwen/Qwen3-4B
• Dataset: centinela-sentiment-es@1
• Formats: safetensors, gguf
• Train config hash: sha256:REPLACE_WITH_REAL_TRAIN_CONFIG_HASH
• Revision SHA: REPLACE_WITH_REAL_HF_REVISION_SHA

Serving

Served generatively via Text Generation Inference (TGI) on an OpenAI-compatible /v1/chat/completions endpoint. Because generation yields no per-label probability, the provider returns SentimentResult.score = None.

How to use

Bind an alias to a live endpoint in config/centinela/bindings.yaml:

apiVersion: astromesh/v1
kind: CentinelaBindings
metadata:
  name: default
spec:
  bindings:
    - model: centinela-sentiment
      alias: prod
      endpoint: https://your-endpoint.endpoints.huggingface.cloud

Compile the bindings into a routable provider with astromeshctl centinela reconcile. Agents then call the typed facade classify(text) -> SentimentResult (exposed as an ADK tool), or route to the model through CentinelaProvider like any other provider.