simpo-training
Pass
Audited by Gen Agent Trust Hub on Apr 4, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: The skill provides documentation and instructional content for machine learning training workflows. No malicious patterns were detected.
- [EXTERNAL_DOWNLOADS]: Instructions include cloning the alignment handbook from Hugging Face's official repository and installing standard Python packages such as torch, transformers, and flash-attn. These are well-known and trusted resources for LLM development.
- [COMMAND_EXECUTION]: The documentation provides shell commands for creating virtual environments with conda, installing packages via pip, and launching training jobs using the accelerate CLI. These are standard practices for the described purpose and are intended for user execution.
- [DATA_EXFILTRATION]: No patterns of sensitive data exposure or exfiltration were found. The skill mentions using datasets from the Hugging Face Hub and interacting with the OpenAI API, which are common and legitimate activities in machine learning workflows.
- [PROMPT_INJECTION]: No instructions attempting to override agent behavior, bypass safety guidelines, or extract system prompts were detected.
Audit Metadata