The Agent Skills Directory

[SAFE]: The skill provides templates and instructions for supervised fine-tuning and reinforcement learning using the TRL and Unsloth libraries on Hugging Face's managed infrastructure.
[COMMAND_EXECUTION]: The scripts/convert_to_gguf.py script uses subprocess.run to perform model quantization. It installs necessary build tools (cmake, build-essential) and compiles the llama-quantize binary from the official llama.cpp repository. These actions are essential for the model conversion process.
[EXTERNAL_DOWNLOADS]: The skill references external scripts and repositories from trusted domains, including huggingface.co and github.com/huggingface. It also clones the well-known llama.cpp project for quantization purposes. These downloads are within the expected scope of a model training and deployment toolkit.
[REMOTE_CODE_EXECUTION]: The skill includes model loading code that uses trust_remote_code=True. This is a standard requirement for loading many transformer-based models with custom logic on Hugging Face. While this allows execution of code from the model repository, it is a standard practice in the machine learning ecosystem and is necessary for the skill's functionality.
[CREDENTIALS_UNSAFE]: Security best practices for secret management are followed by using the $HF_TOKEN environment variable placeholder for Hugging Face authentication, ensuring that sensitive tokens are not hardcoded.

hugging-face-model-trainer