A structurally filtered subset of lambda/hermes-agent-reasoning-traces, pruned from 7,646 to 3,679 rows using automated quality analysis targeting reasoning depth, structural integrity, and tool-call validity.
Most agentic datasets teach models what tool to call but not how to reason about tool selection. The difference matters in production: an agent that dispatches tools without… See the full description on the dataset page:
https://huggingface.co/datasets/ansulev/hermes-agent-traces-filtered.
