docs/evaluation/model_run_workflow.md

Model Run Workflow

This workflow prepares OCR/VLM model runs for v0.2.0 and compares finished prediction files.

Current Local Availability

The development environment may not include OCR or VLM runtimes. The benchmark therefore separates model execution from evaluation:

1. Create a run pack with image paths, output templates, and a shared prompt.
2. Run a model manually or in another environment.
3. Save one prediction JSON file per receipt.
4. Run examples/evaluate_v020_baseline.py.
5. Aggregate multiple model reports with examples/compare_v020_baselines.py.

Prepare A Run Pack

python .\examples\prepare_v020_baseline_run_pack.py `
  --data-root "..\hf_dataset_upload\data\v0.2.0" `
  --model-id "qwen-vl-manual"

The run pack contains:

image_manifest.csv
prompts/v020_receipt_extraction_prompt.md
prediction_templates/
predictions/<model-id>/
README.md

Evaluate A Completed Model Directory

python .\examples\evaluate_v020_baseline.py `
  --data-root "..\hf_dataset_upload\data\v0.2.0" `
  --prediction-dir ".\05_generation\baseline_runs\<run-pack>\predictions\qwen-vl-manual" `
  --output-dir ".\05_generation\generated_reports\baseline_eval_qwen_vl_manual" `
  --fail-on-missing

Compare Multiple Models

python .\examples\compare_v020_baselines.py `
  --reports-root ".\05_generation\generated_reports" `
  --output-dir ".\05_generation\generated_reports\baseline_comparison_latest"

This writes:

baseline_comparison_summary.json
baseline_comparison_summary.csv
baseline_comparison_summary.md

Recommended First Real Baselines

• qwen-vl-*: strong VLM baseline if a local or cloud runtime is available.
• internvl-*: second VLM baseline for structured extraction comparison.
• paddleocr-*: OCR-specialized baseline, useful even if structured JSON is created by a post-processing step.

Do not compare model scores unless the prompt, image variant, prediction schema, and dataset version are recorded.