BIA-MISTRAL-7B-SACHI (merged, bf16)

French ↔ Mooré translation model from BurkimbIA, a non-profit association in Ouagadougou building NLP tools for the languages of Burkina Faso. Mooré (ISO 639-3 mos) is spoken by roughly 8 million people and has very little digital text, which shapes every design choice here.

This is the merged bf16 checkpoint: the one to serve with vLLM.

Which repository do I want?

The SaChi Mistral fine-tune is published three times, from the same training run:

_merged and _4bit are two saves of the same run, five hours apart: every tokenizer file is byte-identical between them, only config.json differs by the quantization_config block bitsandbytes adds.

Do not serve _4bit with vLLM. vLLM handles pre-quantized bitsandbytes checkpoints poorly; use this repository instead and let the server do its own memory management.

Access

This repository is gated: request access above, then pass an approved token to whatever loads it (HF_TOKEN in the environment, --hf-token for vllm serve, token= for from_pretrained). Without it the download fails with 401 Unauthorized on config.json, and a server such as vLLM exits during startup with no other diagnostic.

Prompt format

The model was fine-tuned on a raw completion prompt, not on a chat template. Sending it through a /chat/completions route wraps it in a template it has never seen and degrades the output. Use the completion route with this exact string:

<s>

You are an expert Moore translator. Translate the provided {SRC} text to {TGT}.
The Moore alphabet is: 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.
Based on source language ({SRC}), provide the {TGT} text.
[INST]
### {SRC}:
{TEXT}
[/INST]

### {TGT}:

{SRC} and {TGT} are the English words French and Moore, in either direction. The leading blank line and the <s> are part of the training format; keep them.

Usage with vLLM

Serve it:

vllm serve burkimbia/BIA-MISTRAL-7B-SACHI_merged \
  --served-model-name bia-mistral \
  --dtype bfloat16 \
  --max-model-len 2048 \
  --gpu-memory-utilization 0.92 \
  --enable-prefix-caching

--enable-prefix-caching is worth setting: the prompt above has a fixed ~120-token header shared by every request, so the cache pays for itself immediately (measured hit rate around 50% on mixed traffic).

Call it through the OpenAI-compatible completions route:

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")

PROMPT = """
<s>

You are an expert Moore translator. Translate the provided {src} text to {tgt}.
The Moore alphabet is: 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.
Based on source language ({src}), provide the {tgt} text.
[INST]
### {src}:
{text}
[/INST]

### {tgt}:
"""

def translate(texts, src="French", tgt="Moore"):
    if isinstance(texts, str):
        texts = [texts]
    resp = client.completions.create(
        model="bia-mistral",
        prompt=[PROMPT.format(src=src, tgt=tgt, text=t) for t in texts],
        temperature=0,          # the model was trained for greedy decoding
        max_tokens=512,
    )
    return [c.text.strip() for c in sorted(resp.choices, key=lambda c: c.index)]

print(translate("Le marché est fermé aujourd'hui."))
# ['Raagã pagame rũndã.']

Pass a list of prompts to translate a batch in one call: vLLM runs them together and returns choices in request order. Measured on an RTX 4090, three sentences come back in the same wall-clock time as one.

Decoding: temperature=0. The model is a translator, not a chat assistant, and sampling only adds noise.

Usage with transformers

If you would rather run it directly, prefer _4bit on a smaller GPU. With this repository:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "burkimbia/BIA-MISTRAL-7B-SACHI_merged"
tok = AutoTokenizer.from_pretrained(repo)
tok.padding_side, tok.pad_token = "left", tok.eos_token
model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype=torch.bfloat16, device_map="auto")

inputs = tok(PROMPT.format(src="French", tgt="Moore", text="L'eau est froide."), return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=256, do_sample=False)
print(tok.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=True).strip())

Deployment on RunPod Serverless

Runs unchanged on the stock runpod/worker-v1-vllm image, no custom handler:

MODEL_NAME                        burkimbia/BIA-MISTRAL-7B-SACHI_merged
DTYPE                             bfloat16
MAX_MODEL_LEN                     2048
GPU_MEMORY_UTILIZATION            0.92
ENABLE_PREFIX_CACHING             true
OPENAI_SERVED_MODEL_NAME_OVERRIDE bia-mistral

Needs a 24 GB card or larger (14.5 GB of weights plus KV cache). Endpoint route: https://api.runpod.ai/v2/<endpoint-id>/openai/v1.

Hardware

Related models

• burkimbia/BIA-NLLB-600M-11E — much smaller seq2seq translator, better latency, lower quality on long sentences
• burkimbia/BIA-WHISPER-LARGE-SACHI_V2 — Mooré speech recognition
• burkimbia/BIA-SPARKTTS-V4 — Mooré text to speech

Benchmark and leaderboard: burkimbia/mt-benchmark-public, scored on BLEU, METEOR and chrF across five themes (administrative, daily life, health, religious, arts and technology).

Limitations

• Low-resource. Mooré has little digital text. Output can be disfluent or wrong, and degrades on long, technical or out-of-domain sentences.
• Tone is not written in Mooré orthography, so homographs are frequent and the model can pick the wrong sense.
• The model follows the prompt above closely; changing its wording, the alphabet line or the ### Moore: marker changes the output quality.
• Outputs should be reviewed by a Mooré speaker before any consequential use, in particular for administrative, legal or medical content.

Citation

@misc{burkimbia_sachi_mistral,
  title  = {BIA-MISTRAL-7B-SACHI: French-Moore translation},
  author = {BurkimbIA},
  year   = {2025},
  url    = {https://huggingface.co/burkimbia/BIA-MISTRAL-7B-SACHI_merged}
}