The Agent Skills Directory

[SAFE]: The skill follows professional standards for AI model training. It uses environment variables (e.g., WANDB_API_KEY, TINKER_API_KEY) for secure configuration of external services and provides clear I/O contracts for training scripts.
[INDIRECT_PROMPT_INJECTION]: The skill processes external datasets via EVO_DATASET. While this is a potential attack surface, the skill includes a dedicated reference file (references/false-progress.md) that explicitly warns against and provides detection methods for data contamination, transitive leakage, and reverse-engineering the verifier. This proactive approach significantly mitigates the risk of indirect injection during the training phase.
[EXTERNAL_DOWNLOADS]: The skill references established technology services and academic resources, including Weights & Biases, Hugging Face, and arXiv. These references are used for logging, research, and model hosting in a manner consistent with industry standards.

finetuning