simpo-training
Pass
Audited by Gen Agent Trust Hub on Sep 5, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: Clones the 'alignment-handbook' repository from Hugging Face's official GitHub account.
- [EXTERNAL_DOWNLOADS]: References models and datasets from well-known services including Hugging Face, Meta, Mistral AI, Anthropic, and Nvidia.
- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted preference data (prompts and completions) from external datasets for model training, creating an attack surface where malicious instructions in the data could influence model behavior. 1. Ingestion points: Data loaded via dataset_mixer from external or local files. 2. Boundary markers: No explicit markers mentioned to prevent instruction following during training. 3. Capability inventory: Uses accelerate launch for training and writes output to local storage. 4. Sanitization: Suggests quality filtering but lacks specific sanitization for malicious instructions.
Audit Metadata