The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill references and includes provenance for content from well-known technology organizations and open-source projects, including Hugging Face, PyTorch, NVIDIA, vLLM, DeepSpeed, and EleutherAI. These references are documented neutrally and used for technical guidance.
[COMMAND_EXECUTION]: The skill provides utility scripts such as scripts/vram_ledger.py and scripts/dataset_overlap.py designed for local execution via the uv run command. These scripts perform VRAM estimation and data leakage detection respectively and do not perform suspicious operations.
[INDIRECT_PROMPT_INJECTION]: The skill describes workflows for processing external datasets in JSONL format, which represents a potential attack surface. The skill proactively mitigates this risk by providing rules for loss masking, template verification, and decontaminating training data from evaluation benchmarks.
[CREDENTIALS_UNSAFE]: The file evals/files/api/classify.py contains a placeholder string sk_live_REDACTED. As this is explicitly identified as redacted and located within an evaluation test suite, it does not constitute a sensitive credential exposure.

ml-training