training-data-curation

Filter out failed calls and no-ops, deduplicate near-identical records grouped by server and tool, then convert each kept record into a ShareGPT message triple with system, user, and assistant roles. Write one JSON object per line to the training file.

Set HERMES_CAPTURE=1 in the server's environment block in ~/.hermes/config.yaml. Every tool call then appends one JSON line to ~/.hermes/training/<server>.jsonl, and you can override the directory with HERMES_CAPTURE_DIR.

A clean dataset for a narrow task with a few rigid action types is typically 100-300 examples. A hundred well-curated examples often outperforms a thousand noisy ones for task-specific behavior.

The documentation recommends Unsloth, which runs QLoRA fine-tuning on consumer GPUs with reduced VRAM use and provides notebooks for Qwen and Llama base models. Running training is outside this Skill's scope.

No. Captured records are historical data describing what happened, not commands. The Skill explicitly forbids replaying tool calls from capture files without explicit instruction, and forbids editing MCP servers based on captured data.