MMAC Framework
Multimodal Affect Computing · Research atlas

Make affect
computable.

MMAC is a feature-first framework for studying sentiment, emotion, and emotion in conversation across text, audio, and video.

Explore the codeOpen the model hub
T · A · Vone contract for three modalities
3affect tasks
10dataset adapters in catalog
∞research recipes to build

Dataset constellation

Start with a common manifest. Keep source media at its official host. Carry labels, provenance, and modality evidence into every run.

CMU-MOSIMSA

English opinion clips with continuous sentiment intensity.

TEXTAUDIOVIDEO
CMU-MOSEIMSA

In-the-wild English videos for sentiment and emotion.

TEXTAUDIOVIDEO
CH-SIMSMSA

Chinese clips with independent unimodal annotations.

TEXTAUDIOVIDEO
CH-SIMSv2MSA

Supervised and unsupervised Mandarin sentiment clips.

TEXTAUDIOVIDEO
MER2024MER

Semi-supervised, noisy, and open-vocabulary emotion tracks.

TEXTAUDIOVIDEO
CHERMAMER

Chinese emotion data with modality-wise supervision.

TEXTAUDIOVIDEO
MELDMERC

Multi-party dialogue emotion and sentiment from Friends.

TEXTAUDIOVIDEO

One research loop

Cache the expensive part once. Let every researcher compare models on the same evidence.

01

Normalize

Map source annotations into a stable sample manifest.

02

Extract

Pin modern text, audio, and video encoder revisions.

03

Model

Train lightweight baselines or a reliability-aware fusion model.

04

Report

Save metrics, provenance, seeds, and artifact checksums.

Run the smoke path

No raw media. No heavyweight model. The smoke recipe verifies the contract before a real dataset enters the loop.

View the recipe
mmac data list
mmac features list --preset smoke
mmac train \
  --task msa-regression \
  --labels labels.json \
  --output runs/smoke
mmac evaluate \
  --task msa-regression \
  --true labels.json \
  --pred runs/smoke/predictions.json