wangxiangren/T-MED

T-MED Dataset

T-MED is a teacher multimodal sentiment analysis dataset release (accepted to AAAI2026). This release provides transcripts (text), extracted RoBERTa/HuBERT features, instructional information (subject & stage), and sentiment labels (8 classes).

Download

• Baidu Netdisk (CN mirror): v1: https://pan.baidu.com/s/1viLRTNDrDU21AZVxfBwoww?pwd=2111 // v2： https://huggingface.co/datasets/wangxiangren/T-MED/tree/main
  • Extraction code: 2111
• Version: v1.0 and v2.0

What is Included in This Release This release provides PKL files containing:

• transcripts (text)
• extracted RoBERTa text features
• extracted HuBERT audio features
• instructional information (subject & stage)
• sentiment labels (8 classes)

Each sample contains:

• id: unique sample identifier
• text: transcript text
• roberta_feature: RoBERTa-based text feature
• hubert_feature: HuBERT-based audio feature
• subject: instructional subject
• stage: educational stage / grade level
• label: sentiment label (8 classes)

Raw Audio/Video (By Request) If you need the raw audio/video for further research, please contact the authors by email ( wangxr@mail.imu.edu.cn ).

Suggested information to include in your email:

• Name & affiliation
• Research purpose / intended use
• Which subset (or how much data) you need

Dataset License Dataset License: CC BY 4.0 (recommended) or CC BY-NC 4.0 — see DATA_LICENSE.txt.

Citation If the AAAI proceedings version is available, please cite the official AAAI publication. Otherwise, please cite the arXiv preprint.

(Versioning Notes

The dataset is released in two versions with different feature granularities.

• v1.0 provides pooled, sample-level representations .
• v2.0 provides time-step-level representations with substantially larger storage requirements and is therefore distributed via an external link.)