Train Vela Safety and Hazard

Adapt Vela Safety and Hazard to your application's content policy. Safety predicts safe or unsafe. Hazard identifies the categories of risk, so you can choose a suitable response. Prompt injection and jailbreak detection use the separate Guard model.

To use the published models, follow Safety models. This guide covers training your own compatible checkpoints.

Define the outputs

Both models use the Vela Encoder foundation and a sequence-classification head.

Model	Output	Training objective
Safety	Two-class softmax: safe, unsafe	Cross-entropy
Hazard	Twelve independent sigmoid scores	Masked binary cross-entropy

Hazard uses this ordered label set:

violence, criminal_activity, sexual_content, child_exploitation,
hate, harassment_abuse, regulated_substances, weapons,
self_harm, privacy, specialized_advice, misinformation

A request may have multiple hazards or none. Keep safe negative examples in Hazard training. If a category has not been reviewed, mark it unknown with label_mask: 0 for that category instead of treating it as absent.

Prepare training and evaluation data

Start with the Safety and Hazard data builders for AEGIS and CultureGuard, or use your own labeled requests. Review source labels against your policy: mentioning a sensitive topic, quoting a threat, and requesting harmful action need different judgments.

Include safe educational, preventive, and supportive requests alongside unsafe examples. Keep related documents, conversations, and translations in one split. Prepare training, development, and final test partitions separately.

Each task needs a contract.json with its label maps. Safety rows contain a single label; Hazard rows contain ordered targets and label_mask arrays. The recipe reference documents the full formats and reviewed-data admission command.

Train from Vela Encoder

Use an isolated environment with Transformers 4.57.6 and a platform-appropriate PyTorch build. The training reference lists dependencies. ROCm training uses PyTorch's cuda device API.

Download a fixed Vela Encoder revision to /models/vela-base and set VELA_BASE_REVISION to that revision. The following commands assume your contract and data files are ready.

Train Safety:

python -m src.training.model_classifier.sequence_repair.train \
  --method full --fresh-head \
  --base /models/vela-base --base-id vllm-sr/Vela-1.0-Encoder-307M \
  --base-revision "${VELA_BASE_REVISION:?Set the downloaded revision}" \
  --contract /data/safety/contract.json \
  --train /data/safety/train.jsonl --dev /data/safety/dev.jsonl \
  --output /data/safety/run \
  --steps 600 --batch-size 4 --accumulate 4 \
  --max-length 32768 --microbatch-token-budget 32768 \
  --learning-rate 0.00001 --head-learning-rate 0.0001 \
  --eval-every 200 --evaluation-dtype float32 \
  --selection binary-fp-budget-recall --positive-label unsafe \
  --selection-false-positive-budget 0.1

Train Hazard with the multi-label trainer:

python -m src.training.model_classifier.safety_classifier.train_vela_hazard \
  --method full --fresh-head \
  --base /models/vela-base --base-id vllm-sr/Vela-1.0-Encoder-307M \
  --base-revision "${VELA_BASE_REVISION:?Set the downloaded revision}" \
  --contract /data/hazard/contract.json \
  --train /data/hazard/train.jsonl --dev /data/hazard/dev.jsonl \
  --output /data/hazard/run \
  --steps 600 --batch-size 4 --accumulate 4 \
  --max-length 32768 --microbatch-token-budget 32768 \
  --learning-rate 0.00001 --head-learning-rate 0.0001 \
  --eval-every 200 --evaluation-dtype float32 \
  --selection fp-budget-macro-f1 --selection-false-positive-budget 0.05

These are example budgets and learning rates. Choose the allowed false-positive rate for your application before selecting a checkpoint. Full training saves complete best-model and last-model directories. To continue a compatible Vela task checkpoint, supply it as the base and omit --fresh-head.

Evaluate risks and long inputs

For Safety, measure unsafe recall and false positives on safe requests. For Hazard, measure each category's precision and recall, average precision, and how often safe requests trigger any category.

Choose thresholds on development data, then keep them fixed for the final test. Report languages and categories separately so a large source cannot hide a weak one.

Test short requests alongside 8K, 16K, and 32K documents. Place relevant content at different positions and include benign quotations. The input budget includes special tokens; oversize training rows are reported, and evaluation rejects overflow. Full-context inference and window scanning need separate evaluation because they expose different context to the model.

Export and connect the models

Use the sequence exporter with --method full, the selected checkpoint, and --runtime-task safety or --runtime-task hazard. It preserves the trained label order and writes the weights, tokenizer, and runtime mappings. An ONNX deployment requires graphs exported from those same weights.

Configure the models through local bindings or a supported external service. The Safety signal guide explains how to use a binary risk rule or a category condition. In a category-specific Safety rule, the router runs Hazard after Safety passes its threshold.

Use route preview to verify scores, decisions, errors, and latency with your actual serving configuration.

Earlier Safety models

The earlier mmBERT Safety adapters and nine-class Hazard head remain in the legacy training workflow. Their label contracts differ from Vela's twelve independent Hazard categories.