Hermes · Batch Processing

Generate training data overnight. Or eval a model. Same primitive.

batch_runner.py runs the full agent, with tools, across hundreds or thousands of prompts in parallel, producing structured trajectory data, primarily for fine-tuning data or model evaluation. The input is a JSONL file, one {"prompt": ...} per line (optional per-entry image / cwd). Run with --dataset_file --batch_size --run_name --model --num_workers; each prompt gets its own isolated session. Output lands in data/<run_name>/, headlined by trajectories.jsonl in ShareGPT format (from/value) with per-trajectory tool stats. It is built for long runs: checkpointing plus a content-based --resume that retries failures, automatic quality filtering (no-reasoning and hallucinated-tool samples dropped), and toolset distributions so trajectories cover diverse tool combinations.