EmoMosaic model weights

Deployment weights for RM-Net with BERT LoRA text experts and availability-gated fusion.

• Code and installation: https://github.com/rudykon/EmoMosaic
• Existing data release: https://huggingface.co/datasets/config-h/EmoMosaic

This model repository contains five RM-Net seed checkpoints, one legacy single checkpoint, five LoRA adapters, train-only normalization, frozen fusion settings, and SHA-256 manifests. It contains no raw data, manuscripts or experiment outputs. The checkpoints are custom PyTorch models; use the linked project code, rather than an AutoModel call on this repository. Original checkpoint bytes are preserved.

From a clone of the code repository, run:

make download-weights
make download-models
make download-data
make predict

The downloader restores rmnet/ and rmnet_bert_lora/ under code/models/. The pinned BERT backbone is obtained separately from its original repository; third-party backbone weights are not redistributed here. The alignment downloader also obtains its independent CTC and face-landmark models from their original sources.

The inputs are aligned text, audio and vision features, with explicit observation masks. Outputs include sentiment intensity and negative/neutral/positive probabilities. No new training or performance evaluation was performed for this upload. The historical held-out split was inspected during development and is not an independent blind test. Existing data and upstream components retain their own terms.