---
language:
- fr
- mos
pipeline_tag: text-generation
library_name: transformers
tags:
- moore
- mossi
- burkina-faso
- low-resource
- translation
- instruction-tuned
pretty_name: Tengsoaba 1.7B
---

# Tengsoaba 1.7B

`tengsoaba-1.7b` is a compact instruction-tuned model for **French ↔ Mooré**
(`mos`), built by [BurkimbIA](https://huggingface.co/burkimbia). It is trained to
follow a small set of structured tasks on Mooré, a low-resource language of
Burkina Faso.

A 4B version of the same model, same seven tasks, is published at
[`burkimbia/tengsoaba-4b`](https://huggingface.co/burkimbia/tengsoaba-4b). It scores
**+13.9 chrF** on French → Mooré and is the better choice when the target is Mooré.

The name refers to the *tẽng-soaba*, the master of the land: the custodian of the
earth among the **Nyonyonsé**, the autochthonous people of the region who kept
ritual authority over the land.

## Tasks

The model is multi-task. Each request uses a structured prompt with a `<task>`
tag, an `<instruction>`, and a typed `<input>`.

| Task | Input → Output |
|------|----------------|
| `translate_fr_to_moore` | French → Mooré translation |
| `translate_moore_to_fr` | Mooré → French translation |
| `correct_moore` | Noisy Mooré → corrected Mooré (spelling / OCR / ASR errors) |
| `quality_judgment` | A FR-Mooré pair → `correct` / `incorrect` / `a verifier` + short reason |
| `terminology` | French term → Mooré term (domain-adapted) |
| `standardize_moore` | Mooré → standard orthography |

## Prompt format

The model expects the same structured prompts it was trained on. Example for
French → Mooré:

```
<task>translate_fr_to_moore</task>
<instruction>Traduis le contenu en moore naturel et correct. Produis uniquement la traduction.</instruction>
<alphabet lang="mos">a, ã, b, d, e, ẽ, ɛ, f, g, h, i, ĩ, ɩ, k, l, m, n, o, õ, p, r, s, t, u, ũ, ʋ, v, w, y, z</alphabet>
<input lang="fr">Bonjour, comment allez-vous ?</input>
```

## Usage

```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "burkimbia/tengsoaba-1.7b"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, dtype=torch.float16, device_map="cuda")

prompt = (
    "<task>translate_fr_to_moore</task>\n"
    "<instruction>Traduis le contenu en moore naturel et correct. "
    "Produis uniquement la traduction.</instruction>\n"
    "<alphabet lang=\"mos\">a, ã, b, d, e, ẽ, ɛ, f, g, h, i, ĩ, ɩ, k, l, m, n, o, õ, "
    "p, r, s, t, u, ũ, ʋ, v, w, y, z</alphabet>\n"
    "<input lang=\"fr\">Bonjour, comment allez-vous ?</input>"
)

# apply_chat_template returns a dict on transformers >= 5, so tokenize separately.
text = tok.apply_chat_template(
    [{"role": "user", "content": prompt}],
    tokenize=False, add_generation_prompt=True, enable_thinking=False,
)
ids = tok(text, return_tensors="pt").to(model.device)
out = model.generate(**ids, max_new_tokens=192, num_beams=2, do_sample=False)
print(tok.decode(out[0][ids["input_ids"].shape[1]:], skip_special_tokens=True).strip())
```

**Recommended decoding: `num_beams=2`, no repetition penalty.**

Measured 2026-08-24 on 30 held-out pairs from the `test` split of
`burkimbia/fr_mos_annotated_split_v2`, chrF:

| Decoding | fr→mos | mos→fr | seconds |
|---|---|---|---|
| `num_beams=4`, `repetition_penalty=1.05` | **29.2** | 34.1 | 84 |
| `num_beams=4` | 28.0 | 34.1 | 84 |
| `num_beams=2` | 27.4 | 32.8 | 34 |
| greedy, `repetition_penalty=1.05` | 25.6 | 31.2 | 17 |
| greedy | 24.5 | 31.4 | 25 |
| greedy, `repetition_penalty=1.15` | 22.2 | 32.4 | 19 |

`num_beams=4` scores highest but costs 5x greedy. **`num_beams=2` takes two thirds of the
gain for a third of the extra cost** and is the better default; use 4 when quality matters
more than latency.

**Do not raise `repetition_penalty` above 1.05.** At 1.15, fr→mos drops to 22.2 chrF, the
worst of eight configs. The small gain at 1.05 does **not** transfer to the 4B sibling model,
where any repetition penalty is harmful; measure per checkpoint rather than inheriting a
default.

## Limitations

- **Low-resource.** Mooré has little digital text; the model can produce
  disfluent or wrong output, especially on long or out-of-domain sentences.
- **It invents word forms.** Checked against the 129k-pair training corpus on
  2026-08-24: several outputs contain words with **zero** occurrences in it
  (`yonbẽ`, `wulga`, `marke`, `ne-yẽe`). The output looks like Mooré and uses
  only alphabet-valid characters, which makes invented forms hard to spot
  without a speaker.
- **It translates, it does not converse.** A question asked in Mooré that is not
  a translation request comes back as a restatement of the question, not an
  answer. Use it through the `<task>` prompts above.
- **Tone is not written** in the Mooré orthography, so homographs exist; the
  model can pick the wrong sense.
- `quality_judgment` is a heuristic aid, not a definitive verdict.
- Outputs should be reviewed by a Mooré speaker before any downstream use.