| --- |
| license: apache-2.0 |
| base_model: Helsinki-NLP/opus-mt-yo-es |
| library_name: transformers |
| pipeline_tag: translation |
| tags: [translation, marian, ctranslate2, yoruba, spanish] |
| --- |
| |
| # translate-yo-es — Yoruba to Spanish |
|
|
| Fine-tuned by **Windstorm Labs** from |
| [`Helsinki-NLP/opus-mt-yo-es`](https://huggingface.co/Helsinki-NLP/opus-mt-yo-es). |
|
|
| ## Attribution |
|
|
| Derived from `Helsinki-NLP/opus-mt-yo-es`, copyright the Helsinki-NLP / OPUS-MT project, |
| licensed under **Apache-2.0**. **Modified by Windstorm Labs.** The upstream notice is |
| retained as the licence requires. |
|
|
| ## Contents |
|
|
| | Directory | Contents | |
| |---|---| |
| | `lora/` | **Windstorm Labs fine-tune.** | |
| | `lora-ct2-int8/` | CTranslate2 INT8 build of the fine-tune. Recommended for CPU inference. | |
| | `herm0/` | Upstream weights, retained for reference. | |
| | `herm0-ct2-int8/` | CTranslate2 INT8 build of the upstream weights. | |
|
|
| ## What we did |
|
|
| LoRA fine-tune on GlobalVoices Yoruba–Spanish parallel data — the model's own language pair — |
| merged into the base weights, then quantized to INT8. |
|
|
| | | | |
| |---|---| |
| | Rank / alpha | 8 / 16 | |
| | Learning rate / steps | 2.5e-06 / 50 | |
| | Target modules | `q_proj`, `v_proj` | |
| | Seed | 42 | |
| | Training data | GlobalVoices yo-es, 113 sentence pairs | |
| | Tensors modified | 36 of 259 | |
|
|
| ## Provenance |
|
|
| ``` |
| upstream model.bin sha256 ac8f61aa0c8eb6c36b02c1e656facbe8bd581cc6623a10e5817e223099dbb3dc |
| our model.bin sha256 770efc87c1c46cff851f461666ed8ec1d2e1c69685dfaae5e7dc30412a51acc6 |
| ``` |
|
|
| Verified after INT8 quantization, on the published artifact. |
|
|
| ## Notes |
|
|
| - Yoruba–Spanish is a low-resource pair; quality follows the base model. |
| - No FLORES-200 score is published for this pair because none has been measured. |
| - No human evaluation was performed. |
|
|
