The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill provides instructions to clone the alignment-handbook from Hugging Face's official GitHub repository. This is a well-known and trusted organization in the AI community.
[EXTERNAL_DOWNLOADS]: References and provides configurations for various datasets hosted on the Hugging Face Hub from reputable sources such as Hugging Face, NVIDIA, and Anthropic.
[COMMAND_EXECUTION]: Documentation includes standard shell commands for setting up Python environments via conda and installing necessary machine learning libraries using pip. These are routine developer operations for the stated purpose of model training.
[COMMAND_EXECUTION]: Includes commands to launch training using the accelerate library, which executes local scripts typically found in the cloned alignment-handbook repository.
[PROMPT_INJECTION]: The skill processes preference datasets (e.g., ultrafeedback_binarized) which are ingested from external sources. These datasets constitute an attack surface for indirect prompt injection if they contain malicious instructions.
Ingestion points: External datasets from Hugging Face Hub specified in YAML configurations.
Boundary markers: Not explicitly defined in the provided configuration templates.
Capability inventory: The skill facilitates local filesystem writes for model checkpoints and logs via the output_dir parameter.
Sanitization: No specific content sanitization or filtering of dataset text is described within the skill's instructions.

simpo-training