The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill optionally downloads model checkpoints and vocabulary files from Hugging Face (huggingface.co) using the huggingface_hub library. By default, it targets the official NVIDIA repository nvidia/NV-KERMT-70M-v2.
[COMMAND_EXECUTION]: Orchestrates complex workflows by executing shell commands and Docker containers via subprocess.run and bash scripts. It uses Docker to isolate the training environment and manage GPU access via the NVIDIA Container Toolkit.
[DYNAMIC_EXECUTION]: Loads molecular vocabulary and model checkpoints using pickle and torch.load. The skill implements security mitigations by using a custom _VocabUnpickler that restricts allowed classes to a safe allowlist and utilizes PyTorch's weights_only=True mode with safe_globals to prevent execution of malicious code embedded in artifact files.
[INDIRECT_PROMPT_INJECTION]: The skill represents a potential surface for indirect prompt injection as it ingests and processes untrusted chemical data from user-supplied CSV files.
Ingestion points: Untrusted data enters the context via the --csv and --val-csv inputs processed by scripts/prepare_data.py and scripts/check_data.py.
Boundary markers: The skill does not use explicit boundary markers or instructions to the agent to ignore embedded commands in the processed data.
Capability inventory: The skill has significant capabilities including writing to the file system, executing subprocesses, and running Docker containers.
Sanitization: Partial sanitization is performed via SMILES validation using RDKit in scripts/check_data.py, which ensures inputs are valid chemical notations before further processing.

kermt-continue-pretrain