| | --- |
| | license: apache-2.0 |
| | base_model: |
| | - MHBS-IHB/fish-mt5 |
| | language: |
| | - zh |
| | - la |
| | metrics: |
| | - bleurt |
| | pipeline_tag: text2text-generation |
| | library_name: transformers |
| | tags: |
| | - bilingual |
| | datasets: |
| | - MHBS-IHB/fishmt5 |
| | --- |
| | |
| | # Fine-Tuned mT5 Series for Global Fish Species Dual Latin-Chinese Translation |
| |
|
| | This repository contains fine-tuned versions of the mT5 series models ([mT5-large](https://huggingface.co/google/mt5-large), [mT5-base](https://huggingface.co/google/mt5-base), and [mT5-small](https://huggingface.co/google/mt5-small)) for the specialized task of fish species name translation. These models have been adapted specifically for translating between Chinese and Latin species names. |
| |
|
| | ## Overview |
| |
|
| | By comparing the BLEURT and COMET scores of various models, our experiments demonstrate that the fine-tuned mT5 models significantly outperform general-purpose language models such as DeepSeek-R1, Qwen-Plus, and GLM-Plus. The evaluation covers three translation tasks: |
| |
|
| | - **Chinese-to-Latin** |
| | - **Latin-to-Chinese** |
| | - **Dual Translation** |
| |
|
| | The results, summarized in Table 1 below, indicate that: |
| | - The mT5 series consistently achieves higher scores on both evaluation metrics (BLEURT and COMET) across all translation tasks. |
| | - **fish-mT5-large** stands out, outperforming its counterparts (fish-mT5-base and fish-mT5-small) by achieving BLEURT and COMET scores that are 2 to 3 times higher than those of the traditional models. |
| |
|
| | This substantial performance gap underscores the advantages of the mT5 architecture in handling the nuances of fish species name translation. |
| |
|
| | ## Evaluation Results |
| |
|
| | | Models | Chinese to Latin (BLEURT) | Chinese to Latin (COMET) | Latin to Chinese (BLEURT) | Latin to Chinese (COMET) | Dual Translation (BLEURT) | Dual Translation (COMET) | |
| | |---------------|---------------------------|--------------------------|---------------------------|--------------------------|---------------------------|--------------------------| |
| | | fish-mT5-large| **0.89** | **0.91** | **0.87** | **0.93** | **0.90** | **0.93** | |
| | | fish-mT5-base | 0.80 | 0.87 | 0.77 | 0.87 | 0.80 | 0.88 | |
| | | fish-mT5-small| 0.66 | 0.80 | 0.75 | 0.86 | 0.71 | 0.84 | |
| | | DeepSeek-R1 | 0.44 | 0.66 | 0.56 | 0.74 | 0.45 | 0.67 | |
| | | Qwen-Plus | 0.29 | 0.58 | 0.35 | 0.74 | 0.33 | 0.60 | |
| | | GLM-Plus | 0.26 | 0.55 | 0.33 | 0.60 | 0.30 | 0.58 | |
| |
|
| | *Table 1: BLEURT and COMET scores of six models in dual Latin-Chinese translation for global fish species.* |
| |
|
| | ## Key Takeaways |
| |
|
| | - **Superior Performance:** The fish-mT5 models, particularly fish-mT5-large, demonstrate a clear advantage over traditional models, achieving markedly higher BLEURT and COMET scores. |
| | - **Robustness Across Tasks:** The models perform consistently well in Chinese-to-Latin, Latin-to-Chinese, and dual translation tasks. |
| | - **Specialized Adaptation:** The fine-tuning process has enabled the fish-mT5 architecture to excel in the niche task of fish species name translation, making it a valuable tool for researchers and practitioners in the field. |
| |
|
| | ## Usage |
| |
|
| | We provide an online demo, which can be accessed at [https://huggingface.co/spaces/MHBS-IHB/fishmt5](https://huggingface.co/spaces/MHBS-IHB/fishmt5). |
| |
|
| | --- |
| |
|
| | Feel free to open issues or contribute to the repository if you have suggestions or improvements! |
| | This model card introduces the models, summarizes the evaluation results with the provided table, and highlights the key performance advantages of the fish-mT5-large model for fish species name translation. |
