Local Multimodal Eval

Quick start

Run from the repo root:

chmod +x scripts/run_local_ollama_eval.sh
OLLAMA_MODEL=qwen3-vl:4b OLLAMA_STAGE2_MODEL=qwen3-vl:4b scripts/run_local_ollama_eval.sh

If qwen3-vl:4b is not available locally, set OLLAMA_STAGE2_MODEL to another pulled multimodal model and record that override in the evaluation notes. The operator script now performs a multimodal preflight and should fail fast if the chosen Stage 2 model cannot accept image input.

Workflow

1. Confirm the plan state before tuning. Phase 5 stays open until local multimodal metrics and operational evidence are captured. Phase 6 stays open until final output.csv, final report, and code.zip are produced.