The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill downloads official pretrained model checkpoints from the provider's HuggingFace repository (nvidia/NV-KERMT-70M-v2). This operation requires explicit user consent via an interactive prompt or a specific opt-in flag (--pretrained-release). The download is performed by a local script (fetch_released_model.py) within the container environment.
[COMMAND_EXECUTION]: The skill orchestrates multiple shell commands and scripts to perform system checks, data validation, and model training. These commands are executed using local scripts (kermt_container.sh, run_pretrain_local.py) that are part of the skill's distribution. Training runs are launched in detached Docker containers to support long-running processes without blocking the agent.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a data ingestion surface as it processes user-provided CSV files for training. While this represents a theoretical attack vector for indirect injection, the skill mitigates this risk by employing dedicated validation scripts (check_data.py) and data preparation tools (prepare_data.py) that structure the data into internal shard and feature formats before it reaches the model or runner logic.
[CREDENTIALS_SAFE]: The skill supports integration with Weights & Biases for experiment tracking. It correctly handles credentials by relying on the standard WANDB_API_KEY environment variable and clearly documents that tracking is optional and disabled by default unless the user provides the necessary key.
[DYNAMIC_EXECUTION]: The skill uses python -c to parse system probe results and generates a run.json manifest containing a cmd_replay field. This allows users to reliably reproduce runs using the exact same parameters and environment, providing transparency into the execution logic.

kermt-continue-pretrain