cMSCI — Calibrated Multimodal Semantic Coherence Index

Score how semantically coherent a text / image / audio triple is, with a single calibrated metric.

cMSCI evaluates coherence using pre-aligned embedding spaces (CLIP for text–image, CLAP for text–audio), Gramian-volume geometry, distribution calibration, contrastive margins against hard negatives, a trained cross-space bridge for the image–audio channel, and probabilistic uncertainty weighting.

On a 100-sample human evaluation (5 raters), cMSCI correlates with human coherence judgments at Spearman ρ = 0.785 (p < 1e-6), versus 0.558 for the uncalibrated MSCI baseline.

Install

pip install cmsci

Or the latest development version straight from GitHub:

pip install "cmsci @ git+https://github.com/Pratik25priyanshu20/MultiModal-Coherence-Evaluation-and-Generation.git#subdirectory=cmsci-package"

All trained artifacts (calibration statistics, Ex-MCR projector, cross-space bridge, probabilistic adapters, negative-bank indexes — ~9 MB) are bundled; the CLIP and CLAP backbones download automatically from Hugging Face on first use.

Usage

Python

from cmsci import CoherenceScorer

scorer = CoherenceScorer()
result = scorer.score(
    text="rain falling in a dense forest",
    image="scene.png",     # optional
    audio="rain.wav",      # optional
    domain="nature",       # optional hint for harder negatives
)

print(result.cmsci)           # calibrated coherence score
print(result.msci)            # legacy weighted-cosine baseline
print(result.st_i, result.st_a, result.si_a)   # channel similarities
print(result.variant_scores)  # ablation variants A–F
print(result.uncertainty)     # probabilistic uncertainty estimates

Batch scoring:

results = scorer.score_batch([
    {"text": "waves on a beach", "image": "beach.jpg", "audio": "waves.wav"},
    {"text": "a busy city street", "image": "street.jpg", "audio": "traffic.wav"},
])

Command line

cmsci --text "rain falling in a dense forest" --image scene.png --audio rain.wav
cmsci --text "..." --image scene.png --json     # full result as JSON

How it works

Scores are calibrated against the paper's baseline distribution. For very different domains, recalibration is possible — point CMSCI_ASSETS_DIR at a directory with your own artifacts/cmsci_calibration.json (same layout as the bundled assets).

Configuration

Notes

• Missing assets degrade gracefully: without the bridge the si_a channel is omitted; without calibration, raw uncalibrated scores are returned.
• CPU works fine for scoring; GPU is only beneficial for large batches.

Citation

If you use cMSCI in academic work, please cite the accompanying paper (reference to be added upon publication).

License

MIT