gateremark/kikuyu_translategemma_12b_merged_V2

Kikuyu TranslateGemma-12B: English to Kikuyu Translation

This model, developed by gateremark, is a 12.4 billion parameter language model fine-tuned specifically for English to Kikuyu translation. It is based on Google's TranslateGemma-12B-it and leverages the Unsloth library for 2x faster fine-tuning.

Key Capabilities & Features

• Specialized Translation: Optimized for translating text from English into the Kikuyu language.
• Performance: Achieves a BLEU score of 19.61 on its evaluation dataset.
• Efficient Training: Fine-tuned using LoRA (r=128, alpha=256) with Unsloth and Huggingface's TRL library on 30,430 English-Kikuyu sentence pairs.
• Base Model: Built upon the robust google/translategemma-12b-it architecture.

Intended Use Cases

• Translation Tools: Ideal for integrating into applications requiring English to Kikuyu translation.
• Language Learning: Supports educational platforms for Kikuyu speakers or learners.
• Research: Valuable for studies on low-resource African language NLP.
• Cultural Preservation: Contributes to initiatives focused on preserving the Kikuyu language.

Limitations

• Directional: Currently supports English to Kikuyu translation only; reverse translation was not trained.
• Domain Specificity: Best suited for general text and may face challenges with highly technical or specialized content.
• Dialect Coverage: Trained on standard Kikuyu, potentially limiting performance on specific dialectal variations.