| --- |
| license: apache-2.0 |
| language: |
| - en |
| - ig |
| tags: |
| - translation |
| - nllb |
| - igbo |
| - english |
| - low-resource |
| base_model: facebook/nllb-200-distilled-600M |
| --- |
| |
| # AIN — English ↔ Igbo Translation Model |
|
|
| This model is a fine-tuned version of [facebook/nllb-200-distilled-600M](https://huggingface.co/facebook/nllb-200-distilled-600M), |
| adapted for English -- Igbo translation as part of the **aAIN** project — a speech-to-speech |
| translation system for English and Nigerian languages (Yoruba, Hausa, Igbo). |
|
|
| ## Model description |
|
|
| The base NLLB-200-distilled-600M model was fine-tuned on **27,097 Igbo-English sentence pairs** |
| extracted and cleaned from a bilingual dictionary dataset. Fine-tuning was aimed at improving |
| translation quality specifically for Igbo, a low-resource language with limited high-quality |
| parallel data available for training. |
|
|
| ## Training results |
|
|
| | Metric | Before fine-tuning | After fine-tuning | |
| |---|---|---| |
| | BLEU score | 14.06 | 26.86 | |
|
|
| BLEU improved by +12.8 points on held-out test data — nearly double the baseline score. |
|
|
| ## Intended use |
|
|
| This model is intended for English ↔ Igbo text translation, as part of a larger speech-to-speech |
| pipeline (ASR → translation → TTS). It can be used standalone for text translation via the |
| `transformers` library. |
|
|
| ## Limitations |
|
|
| - **BLEU 26.86 indicates decent but imperfect translation quality.** Expect generally coherent |
| output, but not publication-grade accuracy on every sentence. |
| - Some common conversational phrases (e.g. certain greetings) may occasionally be returned |
| untranslated or copied verbatim from the source. This is a known limitation observed during |
| testing and an area for improvement in future versions. |
| - The training data was sourced from a dictionary dataset, which may bias the model toward more |
| formal or literal phrasing rather than casual/conversational Igbo. |
| - This model does not include Igbo text-to-speech; a separate TTS component is planned for a |
| future release. |
|
|
| ## How to use |
|
|
| ```python |
| from transformers import AutoModelForSeq2SeqLM, AutoTokenizer |
| |
| model_name = "Ntui-Chinua/aAIN" |
| tokenizer = AutoTokenizer.from_pretrained(model_name) |
| model = AutoModelForSeq2SeqLM.from_pretrained(model_name) |
| |
| tokenizer.src_lang = "eng_Latn" |
| inputs = tokenizer("I am going to the market.", return_tensors="pt") |
| forced_bos_token_id = tokenizer.convert_tokens_to_ids("ibo_Latn") |
| |
| generated_tokens = model.generate(**inputs, forced_bos_token_id=forced_bos_token_id, max_length=128) |
| print(tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)[0]) |
| ``` |
|
|
| ## Training data |
|
|
| 27,097 English-Igbo sentence pairs extracted from a bilingual dictionary dataset. |
|
|
| ## License |
|
|
| Apache 2.0 |
|
|
