psychemme/PsycheMME-Training

PsycheMME Training

Training data for the PsycheMME Qwen3.5-4B SFT and GRPO adapters: 3,638 video QA examples, a 2,827-question GRPO pool, video clips, training frames, evidence and source mappings.

Project · Code · Benchmark · SFT Model · GRPO Model

Download

pip install huggingface_hub
hf download psychemme/PsycheMME-Training --repo-type dataset \
  --local-dir PsycheMME-Training
cd PsycheMME-Training
python scripts/prepare.py

The complete download is approximately 59.0 GB. Use Python 3.11 or later. The preparation script extracts the training frames and writes prepared/sft.jsonl and prepared/grpo.jsonl with local media paths. Run it again to reuse extracted frames.

Data

File	Contents
sft/train.jsonl	3,657 SFT examples, including the original assistant targets
grpo/train.jsonl	2,842 single-choice questions in the GRPO candidate pool
grpo/used_ids.txt	The 256 distinct questions used for the reported GRPO updates
media/video/	1,911 video clips and their available audio tracks
media/frames.tar	The 61,152 images used by SFT; GRPO uses a subset of these frames
metadata/episodes.jsonl	Episode IDs, media paths, source links and clip intervals
metadata/media.jsonl	Resolution, frame rate, duration and audio information
metadata/evidence.jsonl	7,060 evidence annotations for all 3,657 SFT questions
metadata/asr.jsonl	The supplied transcript text for 103 questions
metadata/questions.jsonl	Question-to-episode, capability and task-type mappings
metadata/sources.csv	Source video identifiers, links and crop intervals
configs/	SFT and GRPO training settings

JSONL files contain one example per line. SFT examples use id, messages, images and chat_template_kwargs. GRPO examples also include the question, reference solution and metadata used by the reward interface. Frame paths are relative to this repository before preparation; the script resolves them to local files.

media/frames.tar contains the actual 32-frame training inputs. media/video/ provides the corresponding video clips. Evidence text and time fields retain their recorded values.

Training

Setting	SFT	GRPO
Base model	Qwen3.5-4B	Qwen3.5-4B with the SFT adapter
Data	3,657 QA examples	2,842 candidate questions
Recorded training	One epoch, 458 updates	256 distinct questions, 32 updates
Frames per question	32	32
Maximum image dimension	256 px	256 px
LoRA rank / alpha	8 / 32	8 / 32
Learning rate	5e-6	5e-7

The visual encoder and multimodal aligner are frozen. SFT targets contain the answer and one annotated observation. GRPO uses a binary reward for exact answer correctness and response validity, with eight sampled responses per question. Constant-reward and truncated groups are excluded.

The SFT and GRPO repositories provide the corresponding LoRA adapters and loading instructions. Benchmark evaluation is available in the PsycheMME code repository.

Sources

The training examples use EgoCom, Ego4D, YouTube Interview, YouTube Vlog, EasyCom, MERCI+ and CASTLE. Original source links are provided in metadata/sources.csv.

Citation

@misc{hu2026psychemme,
  title={PsycheMME: A Psychology-Informed Benchmark for Social Understanding and Interaction in Videos},
  author={Heran Hu and Yuhang Dong and Zhucun Xue and Kaiwei Chen and Yong Liu and Jiangning Zhang},
  year={2026},
  eprint={2610.28625},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2610.28625}
}