The Agent Skills Directory

[SAFE]: The skill implements a standard machine learning finetuning pipeline for the KERMT molecular encoder. The workflow is well-documented and follows security best practices for this domain.
[EXTERNAL_DOWNLOADS]: The skill provides an option to download a pretrained model from Hugging Face (nvidia/NV-KERMT-70M-v2). Hugging Face is a well-known service, and the resource is maintained by the vendor (NVIDIA). The skill includes a mandatory consent gate, requiring user confirmation before initiating the download.
[CREDENTIALS_UNSAFE]: The documentation identifies the use of WANDB_API_KEY for optional Weights & Biases tracking. This is handled as an external environment variable or user-supplied credential, which is the recommended practice for managing sensitive keys in AI agents.
[INDIRECT_PROMPT_INJECTION]: The skill ingests external data in the form of labeled CSV files. To mitigate risks associated with untrusted data, the workflow includes a dedicated data validation step (check_data.py) and a target confirmation gate where the user must verify the identified target columns before training begins.
[COMMAND_EXECUTION]: The skill executes shell commands and Python scripts within a Docker container (kermt:latest) to ensure environment isolation. These scripts (e.g., prepare_data.py, run_finetune_local.py) are part of the local repository and follow established project patterns.

kermt-finetune