The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the user on executing various shell commands and scripts for environment setup and training, including docker pull, ray job submit, and bash script execution for model training. These are standard operations for a machine learning framework.
[DYNAMIC_EXECUTION]: The framework features several points of dynamic code execution. It provides a load_function(path) utility to dynamically load Python functions from user-defined module paths (e.g., for custom reward models or generation logic). Additionally, it explicitly uses trust_remote_code=True when loading tokenizers and models via Hugging Face, which allows execution of arbitrary code provided within a model repository.
[EXTERNAL_DOWNLOADS]: The documentation includes steps for downloading Docker images from Docker Hub (slimerl/slime:latest), as well as models and datasets from Hugging Face. It also links to official GitHub repositories for THUDM, NVIDIA (Megatron-LM), and the SGLang project.
[INDIRECT_PROMPT_INJECTION]: The skill provides a surface for indirect prompt injection as it processes external training data (JSONL files) which are then interpolated into the LLM training loop. The documentation provides a 'Capability inventory' including subprocess calls and network operations (via SGLang router), which are potential targets for instructions embedded in training data.

slime-user