Results
The Vintern-1B-v2 model was fine-tuned on the maianh511/vi_chart_dataset dataset, showing clear improvements across all metrics compared to the base (pretrain) model.
1. Metric Comparison — Pretrain vs Fine-tuned
| Metric | Pretrain | Fine-tuned | Improvement |
|---|---|---|---|
| BLEU | 0.3054 | 0.4669 | +52.9% |
| METEOR | 0.5510 | 0.6999 | +27.0% |
| ROUGE-1 | 0.6915 | 0.7765 | +12.3% |
| ROUGE-2 | 0.5724 | 0.6739 | +17.7% |
| ROUGE-L | 0.6386 | 0.7321 | +14.6% |
| BERTScore | 0.8542 | 0.9020 | +5.6% |
The fine-tuned model outperforms the base model on every metric, most notably on BLEU (+52.9%) and METEOR (+27.0%) — indicating a substantially better ability to generate Vietnamese answers that closely match the ground truth, both lexically and semantically. BERTScore also improves despite already starting from a high baseline (0.85 → 0.90), showing further gains in semantic quality even though the pretrain model already had a solid foundation.
2. Training Loss Curve
Conclusion
Fine-tuning Vintern-1B-v2 on maianh511/vi_chart_dataset yields consistent improvements across all evaluation metrics, demonstrating the effectiveness of the Vietnamese chart dataset in enhancing the model's ability to understand and answer chart-related questions in Vietnamese.