Skip to content

Privacy And Security

SafeLens configs, reports, and experiments can reference sensitive assets. Keep the repository free of private or large artifacts.

Do not commit:

  • API keys, access tokens, or .env files.
  • Private prompts, private datasets, private annotations, or safety reports that expose user data.
  • Model weights or tokenizer artifacts such as *.pt, *.pth, *.bin, *.safetensors, *.gguf, or *.onnx.
  • Experiment directories such as outputs/, checkpoints/, wandb/, mlruns/, and runs/.

Use local ignored paths for models and generated outputs:

model:
  source: local
  local_dir: ./models/local-causal-lm

output:
  report_path: ./outputs/safety_scan.json

Be careful with trust_remote_code: true. Only enable it for model repositories you trust and have reviewed.