petkopetkov/Qwen3-VL-2B-Instruct-bg

Qwen3-VL-2B-Instruct fine-tuned on subsets of the FineVision dataset translated to Bulgarian language - https://huggingface.co/datasets/petkopetkov/FineVision-bg

Initial results on translated versions of MMMU_val, MMStar and MME:

Benchmark	Metric	Base (Qwen2-VL 2B Instruct)	Finetuned	Δ (Finetuned − Base)	Δ %
MMMU_val_bg	score	0.29	0.2911	+0.0011	+0.38%
MMStar_bg	average	0.0215	0.0074	-0.0141	-65.58%
MMStar_bg	coarse perception	0.0924	0.0000	-0.0924	-100.00%
MMStar_bg	fine-grained perception	0.0112	0.0140	+0.0028	+25.00%
MMStar_bg	instance reasoning	0.0150	0.0067	-0.0083	-55.33%
MMStar_bg	logical reasoning	0.0066	0.0030	-0.0036	-54.55%
MMStar_bg	math	0.0039	0.0000	-0.0039	-100.00%
MMStar_bg	science & technology	0.0000	0.0206	+0.0206	—
MME_bg	mme_cognition_score	411.4286	421.0714	+9.6428	+2.34%
MME_bg	mme_perception_score	1310.1282	1307.4054	-2.7228	-0.21%

Training and evaluation code is available on https://github.com/petkokp/llm_notebooks/tree/main/bulgarian_llms

Uploaded model

• Developed by: petkopetkov
• License: apache-2.0
• Finetuned from model : unsloth/Qwen3-VL-2B-Instruct

This qwen3_vl model was trained 2x faster with Unsloth and Huggingface's TRL library.