The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill facilitates the download of pre-trained models and datasets from the Hugging Face Hub. This is documented neutrally as Hugging Face is a recognized trusted organization. The skill significantly mitigates download risks by requiring users to provide immutable 40-character commit hashes in ST_REVISIONS_JSON and by forcing Hub offline mode during the training process.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data (datasets and models). It addresses this attack surface through a mandatory safety script (training_safety.py) that enforces row-count limits (defaulting to 200, capped at 20,000), disables remote code execution (trust_remote_code=False), and requires the use of SafeTensors for model weights.
[COMMAND_EXECUTION]: The skill utilizes standard command-line tools such as accelerate launch and hf models list for model discovery and distributed training orchestration. These are standard within the intended machine learning workflow and are used within a restricted environment defined by the skill's safety protocols.
[DYNAMIC_EXECUTION]: While model training involves dynamic computation, the skill strictly uses established, well-known libraries (sentence-transformers, transformers, torch). It explicitly prohibits the execution of remote code and untrusted serialized data (pickle weights).
[PRIVILEGE_ESCALATION]: The skill implements least-privilege principles by using os.umask(0o077) in its run preparation script, ensuring that created directories and files are accessible only by the current process owner.

train-sentence-transformers