iamthewalrus67/kulyk-en-uk-grpo
The iamthewalrus67/kulyk-en-uk-grpo is a 0.35 billion parameter English-to-Ukrainian translation model. It is based on the LFM2-350M architecture and has been improved using GRPO (Guardrail-Reinforced Policy Optimization) with calibrated guardrail rewards on WikiMatrix data. This model specializes in high-quality English-to-Ukrainian translation, demonstrating improved BLEU and chrF scores over its baseline on both FLoRes+ devtest and WMT24 en-uk datasets. It is optimized for accurate and contextually appropriate translation between these two languages.
Loading preview...
iamthewalrus67/kulyk-en-uk-grpo: Enhanced English-to-Ukrainian Translation
This model, developed by iamthewalrus67, is a 0.35 billion parameter English-to-Ukrainian translation model. It builds upon the Yehor/kulyk-en-uk (LFM2-350M) baseline and significantly enhances its translation capabilities through GRPO (Guardrail-Reinforced Policy Optimization). The GRPO method incorporates calibrated guardrail rewards during training, leading to more robust and accurate translations.
Key Capabilities & Performance
- Improved Translation Quality: Demonstrates superior performance compared to its baseline model across various metrics.
- On the FLoRes+ devtest (sentence-by-sentence), it achieves a BLEU score of 27.83 and a chrF score of 57.42, outperforming the baseline's 27.28 BLEU and 57.06 chrF.
- For WMT24 en-uk (news domain, out-of-distribution), it shows a BLEU score of 23.23 and a chrF score of 49.80, compared to the baseline's 21.57 BLEU and 48.82 chrF.
- GRPO Training: Utilizes a unique training methodology involving chrF, BLEU, CometKiwi, and five calibrated guardrails as rewards, trained on 132K English-Ukrainian pairs from WikiMatrix.
- Full Fine-tuning: The model underwent a full fine-tune (not LoRA) over 300 steps, with the best checkpoint identified at step 100.
Use Cases
- High-Quality English-to-Ukrainian Translation: Ideal for applications requiring accurate and nuanced translation from English to Ukrainian.
- Research and Development: Can serve as a strong baseline or component for further research in reward-driven machine translation or specific domain adaptation for English-Ukrainian language pairs.