The Agent Skills Directory

[COMMAND_EXECUTION]: The skill orchestrates model training by executing Docker commands (docker build, docker run). It correctly utilizes non-root user mapping (--user $(id -u):$(id -g)) to ensure that artifacts created within containers maintain appropriate host permissions.
[EXTERNAL_DOWNLOADS]: The pipeline downloads model weights, datasets, and reference implementations from trusted sources, primarily HuggingFace and official GitHub repositories for the Transformers library. These are well-known services within the machine learning ecosystem.
[SAFE]: Credential management for HF_TOKEN and WANDB_API_KEY is implemented using standard environment variable passthrough. The skill's instructions explicitly forbid the agent from reading the actual values of these secrets.
[SAFE]: The skill incorporates a mandatory unit testing phase (Phase 4.5) using synthetic data to verify data collation and model forward passes before initiating expensive GPU-based training runs. This reduces the risk of logic errors or resource waste.
[SAFE]: Automated static analysis flags regarding eval() were reviewed and confirmed as false positives; the occurrences were found in pseudocode logic or as part of benign variable names (e.g., eval_results, eval_ds) rather than executable code paths.

tao-finetune-huggingface-model