GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
burkimbia
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
burkimbia
Model Tree
unsloth/mistral-7b-bnb-4bit
Input Modalities
Output Modalities
Supported Functionality
The SaChi Mistral fine-tune is published three times, from the same training run:
| Repository | Contents | Use it for |
|---|---|---|
BIA-MISTRAL-7B-SACHI | LoRA adapter (~100 MB) | applying the fine-tune on top of your own base copy |
BIA-MISTRAL-7B-SACHI_merged | full weights, bf16, 14.5 GB | vLLM, TGI, any fp16/bf16 server |
BIA-MISTRAL-7B-SACHI_4bit (this repo) | full weights, bitsandbytes 4-bit, 7.2 GB | transformers on a 12 GB GPU |
This repository and _merged 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 this repository with vLLM. vLLM handles pre-quantized
bitsandbytes checkpoints poorly; use _merged and let the server manage memory
itself.
This repository is gated: request access above, then pass an approved token
to whatever loads it (HF_TOKEN in the environment, token= for
from_pretrained). Without it the download fails with 401 Unauthorized on
config.json.
The model was fine-tuned on a raw completion prompt, not on a chat template.
Do not wrap it in apply_chat_template. Use this exact string:
text
<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.
python
import torchfrom transformers import AutoModelForCausalLM, AutoTokenizer repo = "burkimbia/BIA-MISTRAL-7B-SACHI_4bit"tok = AutoTokenizer.from_pretrained(repo)tok.padding_side, tok.pad_token = "left", tok.eos_tokenmodel = AutoModelForCausalLM.from_pretrained(repo, device_map="auto") # quantization is in config.json 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", max_new_tokens=256): if isinstance(texts, str): texts = [texts] prompts = [PROMPT.format(src=src, tgt=tgt, text=t) for t in texts] inputs = tok(prompts, padding=True, truncation=True, max_length=1024, return_tensors="pt").to(model.device) with torch.no_grad(): out = model.generate(**inputs, max_new_tokens=max_new_tokens, do_sample=False) gen = out[:, inputs.input_ids.shape[1]:] return [t.strip() for t in tok.batch_decode(gen, skip_special_tokens=True)] print(translate("Le marché est fermé aujourd'hui."))Decoding: do_sample=False. The model is a translator, not a chat assistant,
and sampling only adds noise. bitsandbytes must be installed.
Reverse direction: translate("Raagã pagame rũndã.", src="Moore", tgt="French").
| Precision | Weights | Minimum GPU |
|---|---|---|
| 4-bit (this repo) | 7.2 GB | 12 GB |
bf16 (_merged) | 14.5 GB | 24 GB |
burkimbia/BIA-NLLB-600M-11E — much smaller seq2seq translator, better latency, lower quality on long sentencesburkimbia/BIA-WHISPER-LARGE-SACHI_V2 — Mooré speech recognitionburkimbia/BIA-SPARKTTS-V4 — Mooré text to speechBenchmark and leaderboard: burkimbia/mt-benchmark-public,
scored on BLEU, METEOR and chrF across five themes (administrative, daily life,
health, religious, arts and technology).
### Moore: marker changes the output quality.bibtex
@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_4bit}}<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}:import torchfrom transformers import AutoModelForCausalLM, AutoTokenizer repo = "burkimbia/BIA-MISTRAL-7B-SACHI_4bit"tok = AutoTokenizer.from_pretrained(repo)tok.padding_side, tok.pad_token = "left", tok.eos_tokenmodel = AutoModelForCausalLM.from_pretrained(repo, device_map="auto") # quantization is in config.json 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", max_new_tokens=256): if isinstance(texts, str): texts = [texts] prompts = [PROMPT.format(src=src, tgt=tgt, text=t) for t in texts] inputs = tok(prompts, padding=True, truncation=True, max_length=1024, return_tensors="pt").to(model.device) with torch.no_grad(): out = model.generate(**inputs, max_new_tokens=max_new_tokens, do_sample=False) gen = out[:, inputs.input_ids.shape[1]:] return [t.strip() for t in tok.batch_decode(gen, skip_special_tokens=True)] print(translate("Le marché est fermé aujourd'hui."))@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_4bit}}