---
license: apache-2.0
base_model: Helsinki-NLP/opus-mt-en-mul
language:
- en
- shi
tags:
- translation
- shi
- tatoeba
datasets:
- tatoeba
library_name: transformers
pipeline_tag: translation
---

# English → Tachelhit (opus-mt finetune)

Translates English into Tachelhit. Finetuned by
[Linguardia](https://linguardia.com) from `Helsinki-NLP/opus-mt-en-mul` on the
English↔Tachelhit sentence pairs in [Tatoeba](https://tatoeba.org).

Built to generate course material for a language that no major machine
translator supports.

## Usage

```python
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

tok = AutoTokenizer.from_pretrained("linguardia/opus-mt-en-shi")
model = AutoModelForSeq2SeqLM.from_pretrained("linguardia/opus-mt-en-shi")

batch = tok([">>shi<< I wake up early."], return_tensors="pt")
out = model.generate(**batch, num_beams=4, max_length=64)
print(tok.batch_decode(out, skip_special_tokens=True))
```

The target language is selected by the `>>shi<<` prefix on the source
text. It was added as a **new** token rather than replacing one of the base
model's existing languages, so every language the base already handled still
works.

## Training

| | |
|---|---|
| base | `Helsinki-NLP/opus-mt-en-mul` (77M params) |
| training pairs | 20,399 |
| held out | 1,000 dev, 31 gold |
| epochs | 6.0 |
| batch size | 32 |
| optimiser | Adafactor, lr 5e-5, 500 warmup steps |
| max sequence | 48 tokens (99.9% of source, 99.6% of target) |
| hardware | Apple M4, MPS |

## Evaluation

Held-out Tatoeba (1000 pairs never seen in training):

| metric | score |
|---|---|
| BLEU | 31.0 |
| chrF | 52.8 |

Against 31 human translations of sentences that appear verbatim in the target corpus:

| metric | score |
|---|---|
| BLEU | 37.2 |
| chrF | 59.8 |

Translating the same 2,000-sentence course corpus:

| corpus | rows | distinct | vocab | words/sentence |
|---|---|---|---|---|
| this model | 2000 | 1998 (99.9%) | 2634 | 5.9 |



## Training data and attribution

Sentence pairs come from **Tatoeba**, released under
[CC-BY 2.0 FR](https://creativecommons.org/licenses/by/2.0/fr/), with some
sentences under CC0 1.0. Tatoeba's sentences are written by volunteers, and
this model is a derivative of their work: if you use it, credit Tatoeba and
its contributors.

Pairs were deduplicated on exact (English, Tachelhit) text. No machine
translation was used to create the training data at any point.

## Licence chain

| layer | licence |
|---|---|
| base model `Helsinki-NLP/opus-mt-en-mul` | Apache-2.0 |
| training data (Tatoeba) | CC-BY 2.0 FR / CC0 1.0 |
| these weights | Apache-2.0, with the attribution above |

## Limitations

- Trained on Tatoeba, which is short everyday sentences. Expect it to be
  weakest on long, technical or literary text.
- Neural MT translates each sentence independently, so a term may be rendered
  differently across a corpus. Check consistency if that matters to you.
- Output should be reviewed by a speaker before being taught to learners.