Light-MER Checkpoints

This repository hosts model checkpoints for Light-MER, the official implementation of:

Do We Really Need Multimodal Emotion Language Models Larger Than 1B Parameters?

• Paper: Hugging Face Papers / arXiv
• Code: GAIR-Lab/Light-MER

Checkpoints

File	Description	Status
light-mer-teacher-qwen3-8b.pth	Qwen3-8B teacher checkpoint for Stage 1 SWD-H distillation	Released
stage1-swdh-qwen3-0.6b/checkpoint_000060_loss_1.291.pth	Recommended Qwen3-0.6B student checkpoint after Stage 1 SWD-H distillation	Released
stage1-swdh-qwen3-0.6b/checkpoint_*.pth	Stage 1 SWD-H checkpoint trajectory from epoch 5 to epoch 60	Released
stage1-swdh-qwen3-0.6b/config.yaml	Stage 1 SWD-H training config used for the released checkpoints	Released
stage2-mgrpo-qwen3-0.6b/checkpoint_*.pth	Stage 2 M-GRPO checkpoint sweep for steps 50, 500, ..., 5000	Released
stage2-mgrpo-qwen3-0.6b/config.yaml	Stage 2 M-GRPO training config used for the released checkpoint sweep	Released

Usage

Download the teacher checkpoint and place it under the GitHub repo checkpoint directory:

checkpoints/light-mer-teacher-qwen3-8b.pth

Then run Stage 1 SWD-H distillation with:

CONDA_ENV_NAME=swdh-stage1 \
TEACHER_CKPT=checkpoints/light-mer-teacher-qwen3-8b.pth \
bash scripts/train_stage1_swdh.sh

For Stage 1 inference or evaluation, download one of the released student checkpoints, for example:

checkpoints/light-mer-stage1-swdh-qwen3-0.6b.pth

The recommended Stage 1 checkpoint is:

stage1-swdh-qwen3-0.6b/checkpoint_000060_loss_1.291.pth

For Stage 2 M-GRPO inference and evaluation, download the released sweep directory:

stage2-mgrpo-qwen3-0.6b/

It contains checkpoints for steps 50, 500, 1000, 1500, 2000, 2500, 3000, 3500, 4000, 4500, and 5000. Use this directory as STAGE2_CKPT_ROOT in the GitHub codebase.

License

This checkpoint repository is released under the Apache License 2.0. Please also follow the licenses and usage terms of the external datasets and pretrained models used with Light-MER.

Citation

If you find Light-MER useful, please cite:

@misc{zheng2026lightmer,
  title = {Do We Really Need Multimodal Emotion Language Models Larger Than 1B Parameters?},
  author = {Zheng, Kaiwen and Fu, Junchen and Deng, Wenhao and Han, Hu and Jose, Joemon M. and Ge, Xuri},
  year = {2026},
  eprint = {2607.12787},
  archivePrefix = {arXiv},
  primaryClass = {cs.AI},
  url = {https://arxiv.org/abs/2607.12787},
  note = {Accepted by ACM MM 2026}
}