Raven-1

First-party security models, trained on your human-verified decisions. Every approval and rejection is a labelled training signal — the data flywheel that replaces the baseline LLM.

Training signals↑ 23%

552

human-verified labels

Approved fixes

412

positive examples (SFT)

Corrections

140

rejected / edited

Preference pairs

140

for DPO training

Models

ModelBase (open-weight)Eval vs baselineStatus

raven-1-code-7bdeepseek-coder-6.7b+6.2% vs baselinetraining

raven-1-sec-13bllama-3-8b+3.8% vs baselineevaluating

raven-1-infra-7bmistral-7b—queued

The flywheel

Agent proposes fix→Human approves / rejects→Labelled into corpus→SFT + DPO training (QLoRA, NVIDIA GPUs)→Eval beats baseline→Promote → Triton / NIM serving→Serves the next fix

Served on the NVIDIA stack: TensorRT-LLM + Triton Inference Server, packaged as NVIDIA NIM, trained with QLoRA/DPO and Transformer Engine FP8 on A100/H100.