simpo-training
Pass
Audited by Gen Agent Trust Hub on Oct 5, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: Fetches the alignment-handbook repository from Hugging Face's official GitHub organization to facilitate model training.
- [COMMAND_EXECUTION]: Outlines standard shell commands for creating Conda environments, installing dependencies, and launching distributed training sessions via the accelerate CLI.
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to process external preference datasets from the Hugging Face Hub for model alignment.
- Ingestion points: Datasets are loaded via the
dataset_mixerparameter in configuration YAML files. - Boundary markers: No specific delimiters or safety warnings are included in the instruction templates provided.
- Capability inventory: The training scripts require file system write access to save model checkpoints and logs to the specified
output_dir. - Sanitization: The skill relies on the inherent curation of standard research datasets like UltraFeedback and HelpSteer.
- [SAFE]: The skill provides legitimate technical documentation for researchers, referencing reputable academic papers and established open-source libraries without any malicious patterns.
Audit Metadata