We are pleased to introduce QTuneVL1.5-3B, the latest addition to Reconova AI Lab's series of multimodal large language models. Built upon Qwen2.5-VL-3B, the model's capabilities have been further enhanced through RLVR training using the latest GSPO algorithm.

The model is mainly trained on reasoning datasets, but still maintains proficiency in various general tasks, achieving an overall performance superior to the base model.

Architecture:

ViT: QwenViT
Projector: 2-layer MLP
LLM: Qwen2.5-3B

Evaluation

We evaluate on eight benchmarks specified in the OpenCompass leaderboard using VLMEvalKit, including:

MMBench_TEST_EN/CN_V11, MMStar, MMMU_VAL, MathVista_MINI, HallusionBench, AI2D_TEST, OCRBench, MMVet. The results are shown below:

	Avg	MMBench v1.1	MMStar	MMMU	MathVista	HallusionBench	AI2D	OCRBench	MMVet
Qwen2.5-VL-3B	64.8	77.1	55.3	51.2	60.1	48.6	81.5	83.2	61.4
QTuneVL1-3B	66.1(+1.3)	77.3(+0.2)	57.3(+2.0)	53.6(+2.4)	63.7(+3.6)	49.4(+0.8)	81.3	83.8(0.6)	62.5(+1.1)

The reported results are based on our local implementations and may slightly differ from the official ones.

Copyright

We welcome suggestions to help us improve the QTuneVL. For any query, please contact HanChao Wang: wanghanchao@reconova.com. If you find something interesting, please also feel free to share with us through email or open an issue.

hanchaow/QTuneVL1_5-3B

QTuneVL1.5-3B developed by the Reconova AI Lab (Leader: Jia Baozhi; Team members: Wang Hanchao, Chen Mingmu, Lin Bingqi, et al.) && BDAA-Lab

Introduction

Evaluation

Copyright

hanchaow/
QTuneVL1_5-3B