How a model is made: Centinela

How a model is made: Centinela preview

finanzas español Apache-2.0

Centinela is Nebula’s first model family: Spanish-first models for finance and back-office work in LATAM. It’s the concrete walk-through of everything in the previous pages — the pipeline, the gate, and the catalog contract, applied to one real model.

The model

Centinela-Qwen3-4B (v0.1) is a financial sentiment classifier. Given a Spanish financial sentence, it returns exactly one of three labels:

Contract: positivo · neutral · negativo, enforced with constrain_label — the output is guaranteed to be one of the three, never free text.
Base: QLoRA fine-tune of Qwen/Qwen3-4B.
Dataset: centinela-sentiment-es@1 (finance sentiment, ES).
Formats: safetensors + Q4_K_M GGUF — runs cheap, self-hosted, even on CPU.
License: Apache-2.0.

Its journey through the foundry

flowchart TB
  base["Qwen/Qwen3-4B"] --> train["02_train (QLoRA)"]
  dataset["ES finance dataset"] --> train
  train --> merge["03_merge"] --> quantize["04_quantize"] --> gguf["export/centinela-4b.Q4_K_M.gguf"]
  gguf --> evaluate["05_evaluate: macro_f1 ≥ 0.85 AND beats base?"]
  evaluate -- pass --> push["06_push_to_hub"]
  evaluate -- fail --> abort["abort (no publish)"]
  push --> repo["astromesh/Centinela-Qwen3-4B @ v0.1"]
  push --> catalog["catalog/*.yaml"] --> compile["compile"] --> lock["catalog.lock.json"] --> card["Models card"]

Every step is the real pipeline from The Foundry Pipeline; the gate thresholds are the real configs/eval.yaml; the contract and aliases are the real catalog/centinela.yaml.

Using it

Once published, Centinela is consumable two ways.

Directly, self-hosted

The GGUF build runs anywhere Ollama does — no GPU required:

ollama run hf.co/astromesh/Centinela-Qwen3-4B:Q4_K_M

Through the runtime

An agent routes to it through the GitOps contract — no endpoint or label list copied by hand:

# in an agent's model block
providerRef: centinela-sentiment

The runtime resolves the live endpoint, the positivo/neutral/negativo contract, and auth from the provisioned providers.centinela.yaml. Agents then call the typed facade classify(text) -> SentimentResult (exposed as an ADK tool) or route through CentinelaProvider like any other provider.