Modelpublic

maianh511/qwen2_vl_finetune_lora_viet_chart_vqa

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face

0likes

Model Card

Results — Qwen2-VL LoRA vs Vintern-LoRA

Both Qwen2-VL and Vintern-1B-v2 were fine-tuned with LoRA on the same Vietnamese chart dataset and evaluated on the same 700-sample test set for direct comparison.

Metric Comparison

[image]

Metric	Qwen2-VL LoRA	Vintern-LoRA	Improvement (Vintern vs Qwen2-VL)
BLEU	0.239	0.468	+95.8%
METEOR	0.456	0.703	+54.2%
ROUGE-1	0.613	0.778	+26.9%
ROUGE-2	0.491	0.676	+37.7%
ROUGE-L	0.565	0.735	+30.1%
BERTScore	0.786	0.903	+14.9%

Across all 700 evaluation samples, Vintern-LoRA consistently and substantially outperforms Qwen2-VL LoRA on every metric. The gap is largest on BLEU (+95.8%) and METEOR (+54.2%), suggesting Vintern-LoRA produces Vietnamese answers with far closer lexical and semantic alignment to the ground truth. Even on BERTScore, where both models already score relatively high, Vintern-LoRA maintains a clear lead (0.786 → 0.903).

Conclusion

Under identical fine-tuning setup and evaluation conditions, Vintern-1B-v2 (LoRA) demonstrates significantly stronger performance than Qwen2-VL (LoRA) on the Vietnamese chart QA task, making it the more suitable base model for this use case.