The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The script scripts/generate.py is designed to make network requests to external OpenAI-compatible API endpoints (e.g., Together AI, OpenAI, or local vLLM instances). These requests are used to generate synthetic training data and are configurable via the --base-url argument and OPENAI_BASE_URL environment variable.
[COMMAND_EXECUTION]: The skill provides two Python utility scripts, scripts/generate.py and scripts/curate.py, which are intended to be executed by the user as part of the data preparation pipeline. These scripts use standard Python libraries for networking and file processing.
[DATA_EXFILTRATION]: The generation script transmits prompt instructions and taxonomy metadata to a remote LLM API. This is the primary and intended purpose of the tool to facilitate synthetic data creation. API keys are handled securely via environment variables or command-line flags.
[INDIRECT_PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface through its automated generation pipeline.
Ingestion points: scripts/generate.py reads taxonomy axes from a JSON file and interpolates them into LLM prompts.
Boundary markers: The prompt template uses a basic structured request format but lacks explicit delimiters or instructions to ignore embedded content within the taxonomy variables.
Capability inventory: The pipeline has the capability to perform network operations (API calls) and write files (.jsonl output).
Sanitization: The script performs basic JSON parsing and key extraction but does not sanitize the content of taxonomy values before they are interpolated into the prompt string.

building-finetuning-datasets