{"data":{"kind":"file","path":"README.md","version_id":"bblmtn2zruvau7wzazkul8rd","entry":{"name":"README.md","path":"README.md","is_directory":false,"size":1930,"modified_at":"2026-02-08T13:05:49.627000","content_hash":"b1440d62696c2e3971af2f45682a421e35e245ff36824686135be4f352dab6c4"},"entries":[],"content":"# object-detection-vl\n\n### Overview\n- **Environment ID**: `object-detection-vl`\n- **Short description**: Object detection / bounding box env for VLM models (Qwen VL like)\n- **Tags**: Vision, Bounding box, Object detection, Bbox\n\n### Datasets\n- **Primary dataset(s)**: UlrickBL/elevation-dataset-synthetic-v2. Custom dataset of images with annotated 2D bounding boxes  \n  Format example:\n  ```json\n  [\n    {\"bbox_2d\": [607, 754, 639, 810], \"label\": \"window\"},\n    {\"bbox_2d\": [123, 229, 155, 285], \"label\": \"window\"}\n  ]\n- **Split sizes**: 22 for now (working on synth data)\n\n### Task\n- **Type**: single-turn\n- **Parser**: XMLParser\n- **Rubric overview**:  Evaluates prediction quality based on:\n- Parse correctness\n- Object count accuracy\n- Order-invariant matching (Hungarian algorithm)\n- Bounding box IoU\n- Label correctness\n\n### Quickstart\nRun an evaluation with default settings:\n\n```bash\nuv run vf-eval object-detection-vl\n```\n\nConfigure model and sampling:\n\n```bash\nuv run vf-eval ocr-vl   -m gpt-4.1-mini   -n 20 -r 3 -t 1024 -T 0.7 \n```\n\nNotes:\n- Use `-a` / `--env-args` to pass environment-specific configuration as a JSON object.\n\n### Environment Arguments\n| Arg | Type | Default | Description |\n| --- | ---- | ------- | ----------- |\n| `size` | int | `None` | Limit dataset size |\n| `max_size` | int | `640` | Longest dimension for smart image resize |\n| `iou_weight` | float | `0.7` | Weight assigned to IoU in matching score |\n| `label_weight` | float | `0.3` | Weight assigned to label correctness |\n\n\n### Metrics\n| Metric | Meaning |\n| ------ | ------- |\n| `format_reward` | Whether output successfully parsed into a detection list |\n| `count_reward` | Closeness of predicted vs GT number of objects |\n| `bbox_reward` | IoU-based Hungarian-matched bounding box localization reward |\n| `label_reward` | Hungarian-matched label correctness |\n| `final_reward` | Main scalar reward (weighted sum of all components) |\n\n\n","encoding":"utf-8","truncated":false,"total_bytes":1930},"status":null}