burkimbia

tengsoaba-4b

README

License: apache-2.0

Tasks

Table with columns: Task, Input → Output
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 prompts it was trained on. <alphabet> is emitted for translate_fr_to_moore only: the constraint applies when the output is Mooré.

markdown
<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-4b"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, dtype=torch.bfloat16, 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())

enable_thinking=False is required, not optional. Qwen3 thinking mode is on by default, produces its trace in Chinese, consumes the whole token budget, and degrades answers.

Do not use repetition_penalty on this model. Measured 2026-08-24 on 30 held-out pairs from the test split of burkimbia/fr_mos_annotated_split_v2, chrF:

Table with columns: Decoding, fr→mos, mos→fr, seconds
Decoding	fr→mos	mos→fr	seconds
`num_beams=2`	39.6	33.9	2026
greedy	38.4	32.1	923
greedy, `repetition_penalty=1.05`	34.0	31.2	819
greedy, `repetition_penalty=1.15`

The penalty costs 4.4 chrF at 1.05 and 7.7 at 1.15. This is the opposite of the 1.7B sibling, where 1.05 helps slightly. Decoding settings do not transfer between checkpoints; measure per model.

num_beams=2 buys 1.2 chrF for 2.2x the time. Greedy is a reasonable default here, because a stronger model makes fewer search errors, so beam has less to recover.

How it compares to the 1.7B sibling

Same test set, same prompts, chrF:

Table with columns: 1.7B, 4B, Δ
	1.7B	4B	Δ
fr→mos, greedy	24.5	38.4	+13.9
mos→fr, greedy	31.4	32.1	+0.7

Capacity helps in one direction only. Generating into a low-resource language is the hard part and benefits from parameters; generating into French is already saturated at 1.7B. An averaged score across directions would have shown +7.3 and hidden this.

Cost of the gain: 8 GB of VRAM against 3.4, and roughly 40x the inference time on a GPU that leaves little headroom after the weights.

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. Its 1.7B sibling was checked against the 129k-pair training corpus and produced words with zero occurrences in it. The outputs use only alphabet-valid characters, which makes invented forms hard to spot without a speaker. Not separately verified on this model; assume the same.
It translates, it does not converse. See the note at the top.
Tone is not written in the Mooré orthography, so homographs exist; the model can pick the wrong sense.
Time-of-day greetings are unreliable. "Bonsoir" and "Bonne nuit" can both come back as the morning greeting.
quality_judgment is a heuristic aid, not a definitive verdict.
Outputs should be reviewed by a Mooré speaker before any downstream use.

Training

Supervised fine-tuning on burkimbia/moore-instruct-v2 (466 250 train rows), 2 epochs. No replay of general-domain data was mixed in, and none was needed: the base capabilities measured intact after training.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

burkimbia

Model Tree

Base

this model

Input Modalities

Text

Output Modalities