Omnibin - Comprehensive ML Metrics Report Generator

Binary Classification Report

Upload a CSV with 'y_true' (0/1) and 'y_pred' (0-1 probability) columns.

Regression Report

Upload a CSV with 'y_true' and 'y_pred' columns (continuous values).

Segmentation Report

Upload two NumPy (.npy) files: ground truth mask and prediction mask.

Object Detection Report

Upload a JSON file with the following structure:

{
    "predictions": [
        [{"box": [x1,y1,x2,y2], "score": 0.9}, ...],
        ...
    ],
    "ground_truths": [
        [[x1,y1,x2,y2], ...],
        ...
    ]
}

Text Generation Report (Radiology Reports)

Upload a CSV with reference and candidate columns. Lexical metrics (BLEU / ROUGE / METEOR / BERTScore) run locally. LLM-judge metrics (GREEN, RadFact, CRIMSON) require an API key.

Paper-default judge models: GREEN → StanfordAIMI/GREEN-RadLlama2-7b (Ostmeier et al., 2024); RadFact → Llama-3-70B-Instruct (Bannur et al., 2024); CRIMSON → MedGemmaCRIMSON-4B (Baharoon et al., 2026).

⚠️ DISCLAIMER — API-mode judges are not paper defaults.
The hosted Space runs GREEN and RadFact in API mode: we use each paper's verbatim prompt (GREEN) or system messages (RadFact) but route them to a generic provider you select (OpenAI / Anthropic / Google / OpenRouter / Groq). Scores can differ from published numbers, because the original papers fine-tuned their own judge models.

• GREEN API mode: verbatim prompt + parser, only the judge model changes.
• RadFact API mode: verbatim two-stage system prompts but with zero-shot JSON output (the upstream radfact package teaches output format with 10-shot YAML). Only logical precision/recall/F1; no grounding or spatial scores.
• CRIMSON API mode: verbatim prompt + JSON parser + scoring formula (vendored from rajpurkarlab/CRIMSON), only the judge model changes. We use API mode on this Space because crimson-score pins pandas>=3.0.1 (conflicts with Gradio's pandas<3.0); tracked upstream as rajpurkarlab/CRIMSON#3.
  For exact paper reproduction: pip install omnibin[green] (needs GPU), pip install omnibin[radfact] (needs a separate env due to pydantic 1.x / Gradio conflict), or pip install omnibin[crimson] (separate env until the pandas pin is relaxed).