Release note: the built-in signals default to Vela 2.0 0.3B

Vela 2.0 is public on Hugging Face: vllm-sr/Vela-2.0-0.3B, -0.8B, -4B and -9B, Apache-2.0, with no token needed. The model runtime keeps the revisions it pins, so nothing it loads changes.

What changes by default

With no model configured, the domain, prompt guard, safety, fact check, user feedback, modality, PII and hallucination signals run on one deployment of vllm-sr/Vela-2.0-0.3B, @Vela-2.0-0.3B, in one call per request. They no longer run on eight Vela 1.0 specialists.

• Questions: each signal asks the question the model was trained on for it, with the labels of its Vela 1.0 model, so rules and policies read the answer as before.
• CPU profile: on a CPU the deployment runs max_speed, which gives the same answers to within about 0.00001.
• Input: the model reads up to 8,192 tokens of a request and truncates the rest. The Vela 1.0 Guard and PII specialists scanned up to 32K in windows.
• Modality: a modality_detector with method: classifier and no classifier.model_path now runs the 0.3B. It used to be a configuration error.
• What stays: Hazard, embeddings, multimodal embeddings and the reranker keep their Vela 1.0 models.
• Operator: the operator's prompt_guard.model_id defaults to the 0.3B, and its threshold to 0.75. A SemanticRouter created earlier keeps the values it stored.

The maintainers chose this default against two of its goals

#4639 asked for accuracy level or better on every signal and router latency level or better, both measured on a CPU. The maintainers chose to switch although both fail, accepting the modality and user feedback regressions. Through the Router, on the router signal suite (A/B record):

• Ahead: prompt guard (held-out AUC +0.026) and safety (+0.052).
• Level: PII and hallucination on held-out and fresh files.
• Behind, most: modality (held-out AUC −0.180; the 0.3B misses most requests for a new image) and user feedback (accuracy −0.038 held-out, −0.178 fresh).
• Behind: domain (accuracy −0.037 held-out) and fact check (held-out AUC −0.101).
• CPU latency: on 12 cores a request takes 79 ms at the median against 16 ms on Vela 1.0, about 4.9 times as long, and the Router serves 11.9 against 38.9 requests per second (12.8 against 51.8 at concurrency 16). Every request carries the questions, their options and the 17 PII labels (at least 560 tokens) through one forward. #4668 works on the CPU latency. On a GPU the 0.3B answers in about 7 ms.

Thresholds

The module defaults are recalibrated to the 0.3B's scores so that each keeps the Vela 1.0 specialist's operating point on the suite's dev split: prompt guard 0.5 → 0.75, domain 0.5 → 0.28, PII 0.9 → 0.01, fact check 0.95 → 0.93, user feedback 0.7 → 0.37.

• PII: the model's span head applies its own per-label thresholds before it returns a span, so 0.01 accepts every span it returns.
• Other models: a module that runs any other model and sets no threshold keeps its earlier default.
• Maintained configurations: the recipes and E2E profiles that run the defaults carry the mapped values of their rule thresholds.
• Your own rule thresholds were likely chosen for Vela 1.0: map them with the record's table (prompt guard 0.3–0.9 → 0.74–0.77, PII → 0.01, safety 0.5 → 0.46, modality confidence_threshold 0.7 → 0.51), or restore Vela 1.0.

Restore Vela 1.0

One block brings the specialists back, with their default thresholds:

global:
  model_catalog:
    system:
      safety: models/Vela-1.0-Encoder-307M-Safety
      prompt_guard: models/Vela-1.0-Encoder-307M-Guard
      domain_classifier: models/Vela-1.0-Encoder-307M-Domain
      pii_classifier: models/Vela-1.0-Encoder-307M-PII
      fact_check_classifier: models/Vela-1.0-Encoder-307M-FactCheck
      hallucination_detector: models/Vela-1.0-Encoder-307M-Halu
      feedback_detector: models/Vela-1.0-Encoder-307M-Feedback

One line brings back one signal, and the others stay on the 0.3B:

Signal	Line under global.model_catalog
Domain	system.domain_classifier: models/Vela-1.0-Encoder-307M-Domain
Prompt guard	system.prompt_guard: models/Vela-1.0-Encoder-307M-Guard
Safety	system.safety: models/Vela-1.0-Encoder-307M-Safety
Fact check	system.fact_check_classifier: models/Vela-1.0-Encoder-307M-FactCheck
User feedback	system.feedback_detector: models/Vela-1.0-Encoder-307M-Feedback
PII	system.pii_classifier: models/Vela-1.0-Encoder-307M-PII
Hallucination	system.hallucination_detector: models/Vela-1.0-Encoder-307M-Halu
Modality	modules.modality_detector.classifier.model_path: models/Vela-1.0-Encoder-307M-Modality

A signal moved back takes its Vela 1.0 rule thresholds with it, for example mom-v1's prompt guard 0.5, safety 0.5 and PII 0.7. See Choose a model.