Olala-v1.0-translate

Olala-v1.0-translate is a machine translation model, fine-tuned from Olala-v1.0-mini-7A1B and specialized for financial documents.

  • 18 languages: European languages plus Arabic, Chinese (simplified and traditional), Japanese and Korean.
  • Controllable: follows a terminology mapping you supply, adapts to a document domain (annual reports, fund factsheets, prospectuses, regulatory texts) and preserves casing and inline tags.
  • Best on financial documents: 62.6 BLEU on DOLFIN English to French, 8.6 points ahead of the next model we compared, the best COMET-22 score (88.0) on news and financial text, and within 0.3 BLEU of the best on WMT news.

Model details


Model Description
Olala-v1.0-mini-7A1B-base Pre-trained base model (12.8T tokens)
Olala-v1.0-mini-7A1B-SFT After supervised fine-tuning, before RL
Olala-v1.0-mini-7A1B Final post-trained model: reasoning, chat and tool use
Olala-v1.0-summarise Specialized for document summarization
Olala-v1.0-translate Specialized for machine translation, including financial documents
Property Specification
Checkpoint Fine-tuned from Olala-v1.0-mini-7A1B for translation (merge of SFT, ORPO and GRPO checkpoints)
Parameters 6.8B total, about 1B active per token
Architecture 36 layers: 29 Mamba-3 and 7 attention layers, mixture-of-experts with 256 routed experts (6 active)
Context length 32K tokens (input + output)
Translation fine-tuning and evaluation used inputs up to 8K tokens.
Languages English, French, German, Spanish, Italian, Dutch, Portuguese, Swedish, Danish, Norwegian, Finnish, Polish, Czech, Russian, Arabic, Chinese (simplified and traditional), Japanese, Korean
License Apache 2.0

Inference


The model was trained on a fixed instruction followed by a JSON request, with reasoning disabled. You do not need to build it by hand: the repo ships a translation chat template, additional_chat_templates/translation.jinja, that renders it from the user message and a few chat_template_kwargs. The default chat_template.jinja is the general chat template.

Field Required Value
source_language yes English language name: English, French, German, Spanish, Italian, Dutch, Portuguese, Swedish, Danish, Norwegian, Finnish, Polish, Czech, Russian, Arabic, Chinese, zh-Hans, zh-Hant, Japanese, Korean
target_language yes same list
domain no default General. Trained domains include Annual Report IFRS, Fund Factsheet, Prospectus KIIDs, Regulatory, Asset Management Marketing, Equity Research, Life Insurance, Corporate Actions, Custom
terminology_mapping no default []; list of {"source_term": "...", "target_term": "..."}
source_text no default: the content of the last user message

Decoding: greedy (temperature=0), as in the evaluation.

vLLM: always serve this model with --mamba-ssm-cache-dtype float32. vLLM's default keeps the Mamba recurrent state in bf16, and long generations then degrade into repetition loops.

vLLM quick start (vLLM fork, installed as for Olala-v1.0-mini-7A1B):

vllm serve OVHaiLLM/Olala-v1.0-translate \
    --served-model-name olala-translate \
    --trust-remote-code \
    --chat-template additional_chat_templates/translation.jinja \
    --reasoning-parser olala \
    --mamba-ssm-cache-dtype float32 \
    --max-model-len 32688

--chat-template makes every request on this server a translation request. To keep the general chat template, leave it out and send the content of translation.jinja in each request's chat_template field. --max-model-len must be a multiple of 144.

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")
r = client.chat.completions.create(
    model="olala-translate",
    messages=[{"role": "user", "content": "Le rapport annuel du fonds est publié en mars."}],
    temperature=0,
    max_tokens=2048,
    extra_body={"chat_template_kwargs": {
        "source_language": "French",
        "target_language": "English",
        "domain": "Fund Factsheet",
        "terminology_mapping": [{"source_term": "fonds", "target_term": "fund"}],
    }},
)
print(r.choices[0].message.content)

Transformers quick start:

from huggingface_hub import hf_hub_download
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "OVHaiLLM/Olala-v1.0-translate"
tok = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(repo, trust_remote_code=True, dtype="bfloat16",
                                             device_map="cuda")  # or "cpu" without a GPU (slow)
template = open(hf_hub_download(repo, "additional_chat_templates/translation.jinja")).read()

inputs = tok.apply_chat_template(
    [{"role": "user", "content": "Le rapport annuel du fonds est publié en mars."}],
    chat_template=template, add_generation_prompt=True, return_tensors="pt", return_dict=True,
    source_language="French", target_language="English", domain="Fund Factsheet",
    terminology_mapping=[{"source_term": "fonds", "target_term": "fund"}],
).to(model.device)

out = model.generate(**inputs, max_new_tokens=512, do_sample=False, use_cache=True)
text = tok.decode(out[0, inputs["input_ids"].shape[1]:])
print(text.split("<|content|>")[-1].split("<|channel_end|>")[0])   # the final channel: the translation
Chat template

The repo ships two chat templates:

  • chat_template.jinja: the general Olala 1.0 mini chat template (default), with reasoning in an analysis channel and the answer in a final channel.
  • additional_chat_templates/translation.jinja: the translation template. The client sends only the text to translate as the user message, and the request fields as chat_template_kwargs. The template writes the exact prompt the model was trained on: the translation instruction, then the JSON request (keys in the order source_language, target_language, source_text, terminology_mapping, domain).

The translation template always disables reasoning: the rendered prompt ends with <|im_start|>assistant<|channel_start|>analysis<|content|><|channel_end|>, so the model answers directly in the final channel. With the vLLM reasoning parser, the response content is the translation only.

A request without source_language or target_language fails with translation template: source_language and target_language are required (chat_template_kwargs).

Performance


Benchmarks

Olala-v1.0-translate against open translation and instruction models of up to 7B parameters, on the same data, with greedy decoding:

  • WMT newstest: sentence-level translation quality on the WMT news test sets, 28 directions among English, French, German, Spanish, Italian, Dutch, Portuguese, Swedish, Danish, Norwegian and Finnish. BLEU averaged over the 28 directions.
  • COMET-22 (Unbabel/wmt22-comet-da, times 100) is computed on the full 28-pair set, i.e. the WMT news test sets plus 9 financial domains (fund factsheets, KIIDs, annual reports, regulatory, equity research, ...), not on the news test sets alone.
  • DOLFIN (Nakhlé et al., Findings of NAACL 2025; dataset: OVHaiLLM/dolfin): document-level financial translation. 340 English→French KIID / PRIIPs documents (about 200 words each), each translated in one pass. BLEU.
Model Size WMT newstest
28 directions, BLEU
News + financial
28 pairs, COMET-22
DOLFIN en→fr
financial documents, BLEU
Olala-v1.0-translate7B-A1B32.688.062.6
Gemma 4 E4B4B31.987.454.0
Gemma 3 4B4B28.786.151.9
Qwen3.5 4B4B29.686.551.7
Hy-MT2 7B7B29.186.251.7
Tower+ 2B2B32.987.550.8
TranslateGemma 4B4B26.886.648.6
Hy-MT2 1.8B1.8B25.283.347.5
Qwen3 4B 25074B26.984.746.8
Ministral 3 3B3B20.979.736.7

Olala-v1.0-translate runs with reasoning disabled and the JSON translation request above. Each other model is shown with its best run per benchmark: its native prompt where its model card documents one (TranslateGemma, Hy-MT2, Tower+), otherwise the better of the same JSON request and a plain-text prompt with the same instructions.

Training


Starting from Olala-v1.0-mini-7A1B, the checkpoint is an equal-weight average (model merge) of three stages of the same translation run:

  1. Supervised fine-tuning on multilingual translation data, general and financial, with and without terminology constraints.
  2. Preference optimization (ORPO) on top of the SFT model.
  3. Reinforcement learning (GRPO) from the ORPO model, with metric-based rewards and no LLM judge.

Uses and limitations


Olala-v1.0-translate is built for machine translation, especially of financial documents. The limitations of Olala-v1.0-mini-7A1B apply. In addition:

  • Translation only. Use it with the translation template; for chat, use Olala-v1.0-mini-7A1B.
  • Supported languages only: the languages listed above. Other pairs are not supported.
  • Long documents: fine-tuning and evaluation used inputs up to 8K tokens. Split longer documents.
  • Not a certified translation. Have a qualified translator review legal, regulatory and contractual texts before you publish or rely on them.
  • Terminology: the model follows the mapping you supply but can miss terms. Check key terms in the output.

Before you deploy: read the public summary of training content and our copyright policy.

License


Licence Apache License, Version 2.0 (SPDX: Apache-2.0)
Licensor OVH LLM SAS, trading as OVHai LLM, 2 rue Kellermann, 59100 Roubaix, France
Copyright Copyright 2026 OVH LLM SAS
Release date 19 November 2026
Distribution Open weights
Notices NOTICE: keep it when you redistribute the model or a derivative. It includes the attribution of the Qwen tokenizer (Copyright 2024 Alibaba Cloud, Apache 2.0)
Training content (EU AI Act) Public summary of training content, on the European Commission template (Art. 53(1)(d))

Citation


@techreport{ovhaillm2026olala,
  title       = {Olala v1.0 mini Technical Report},
  author      = {{OVHai LLM Research Team}},
  institution = {OVHcloud},
  year        = {2026},
  url         = {https://huggingface.co/OVHaiLLM/Olala-v1.0-mini-7A1B/blob/main/Olala-v1.0-mini.pdf},
}

To cite the translation model itself:

@misc{olala-v1.0-translate,
  title  = {Olala-v1.0-translate: a hybrid Mamba-3 / MoE model for financial machine translation},
  author = {OVHaiLLM},
  year   = {2026},
  url    = {https://huggingface.co/OVHaiLLM/Olala-v1.0-translate}
}
Downloads last month
69
Safetensors
Model size
7B params
Tensor type
F32
·
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for OVHaiLLM/Olala-v1.0-translate

Collection including OVHaiLLM/Olala-v1.0-translate