mt5-base-nepali-gec-stage2
Description
This model is a refined version of mt5-base-nepali-gec-stage1, further fine-tuned on high-quality data
to improve fluency and accuracy.
Intended Use
- Improved grammatical error correction in Nepali
- Production-ready GEC model
Training Data
- Synthetic dataset (~1M sentences)
- High-quality LLM-generated dataset (~8K sentences)
Training Details
- Base model: mT5-base
- Fine-tuning: Stage 1 + Stage 2 (refinement)
Limitations
- May still struggle with rare or ambiguous errors
- Performance depends on input quality
Example
Input: "उ घर गयो थियोन"
Output: "उ घर गएको थिएन"
