The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides numerous shell commands for the agent to execute, including environment management (conda create), package installation (pip install), and launching distributed training jobs (accelerate launch).
[EXTERNAL_DOWNLOADS]: The instructions require cloning the alignment-handbook repository from Hugging Face and fetching various preference datasets (e.g., UltraFeedback, Anthropic HH-RLHF) from the Hugging Face Hub. Hugging Face is a recognized trusted organization.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external datasets for preference alignment, creating an attack surface where malicious instructions embedded in the data could influence the agent during data processing or training validation.
Ingestion points: Preference datasets loaded via dataset_mixer in SKILL.md and references/datasets.md (e.g., HuggingFaceH4/ultrafeedback_binarized).
Boundary markers: The instructions recommend checking data for invalid values in the 'Agent operating procedure', but do not define strict delimiters or safety-specific sanitization for the prompt content.
Capability inventory: The skill utilizes shell command execution, file system writes (configurations/outputs), and network access (git/pip/Hugging Face).
Sanitization: The skill provides examples of quality-based filtering (length, diversity) in references/datasets.md, but lacks explicit security sanitization to prevent adversarial instructions from affecting the agent's logic.

simpo-training