maianh511/internvl2_1b_finetune_lora_viet_chart_vqa

Results — InternVL-FT vs Vintern-LoRA

InternVL (full fine-tuning) and Vintern-1B-v2 (LoRA) were evaluated on the same Vietnamese chart dataset for direct comparison.

Metric Comparison

Metric	InternVL-FT	Vintern-LoRA	Improvement (Vintern vs InternVL)
BLEU	0.253	0.468	+85.0%
METEOR	0.512	0.703	+37.3%
ROUGE-1	0.621	0.778	+25.3%
ROUGE-2	0.476	0.676	+42.0%
ROUGE-L	0.567	0.735	+29.6%
BERTScore	0.837	0.903	+7.9%

Vintern-LoRA outperforms InternVL-FT across every metric, even though InternVL was fully fine-tuned while Vintern only used LoRA. The largest gaps appear on BLEU (+85.0%) and ROUGE-2 (+42.0%), indicating Vintern-LoRA generates answers with notably better n-gram and phrase-level overlap with the ground truth. On BERTScore, both models score relatively high, but Vintern-LoRA still holds a consistent edge (0.837 → 0.903).

Conclusion

Despite using a lighter-weight LoRA fine-tuning approach, Vintern-1B-v2 achieves stronger results than a fully fine-tuned InternVL model on the Vietnamese chart QA task, highlighting both the efficiency of LoRA and the suitability of Vintern as the base model for this task.