SauerkrautLM-Translator-LFM2.5-1.2B 🥬

A specialized translation model fine-tuned for high-quality any-to-any multilingual translation

Model Description

SauerkrautLM-Translator-LFM2.5-1.2B is a compact yet powerful translation model built on top of LiquidAI/LFM2.5-1.2B-Instruct. Unlike conventional translation models that focus primarily on English↔X language pairs, this model has been trained to handle any-to-any translation across five European languages: English, German, French, Spanish, and Italian.

Bonus: Despite being trained exclusively on these five languages, the model demonstrates surprisingly strong zero-shot translation capabilities from Arabic and Chinese as source languages!

The model excels at:

• 🔄 Any-to-any translation between all supported languages
• 📝 Long-form text translation (documents, articles, etc.)
• 💬 Short text translation (sentences, phrases)
• 🎯 Instruction translation – particularly effective at translating prompts and instructions while preserving their instructional intent

Key Features

Feature	Description
Target Languages	English, German, French, Spanish, Italian
Source Languages	EN, DE, FR, ES, IT + zero-shot: Arabic, Chinese
Translation Directions	All 20 language pair combinations (any-to-any) + AR/ZH → X
Parameters	1.2B
Architecture	LFM (Liquid Foundation Model)
Training Data	3M examples + 30K DPO refinement

Training Details

Dataset Composition

The model was trained on a carefully curated dataset of 3 million examples covering all possible translation directions between the five supported languages. The training data includes:

• Long-form translations: Full documents, articles, and extended text passages
• Short translations: Individual sentences and brief text segments
• Instruction-following examples: Specifically designed to ensure the model translates instructions as requested, rather than attempting to execute the underlying task

This diverse training approach makes the model particularly robust for real-world translation scenarios, including the translation of prompts and system instructions.

Synthetic Data Generation

The training dataset was generated using Qwen/Qwen3-Next-80B-A3B-Instruct-FP8.

DPO Refinement

The model was further refined using Direct Preference Optimization (DPO) on 30,000 carefully selected examples. Preferred responses were selected based on higher quality translations from state-of-the-art models, which significantly boosted translation quality and accuracy.

Evaluation

We conducted a comprehensive evaluation using GPT-4.1 as judge, comparing models head-to-head across all language pairs. This resulted in 96,000 total comparisons, designed to minimize benchmark bias and provide a robust assessment of translation quality.

Summary Dashboard

Overall Win Rate

Head-to-Head Comparison

Performance Radar Chart

Wins vs Losses

Performance by Source Language

Performance by Target Language

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

# Load model and tokenizer
model_id = "VAGOsolutions/SauerkrautLM-Translator-LFM2.5-1.2B"
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    torch_dtype="bfloat16",
    trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)

# Translation parameters
text = "Cuál es equivalente a una bombilla de 100w en led"
target_language = "German"  # English, German, Italian, Spanish, French

# Generate translation
input_ids = tokenizer.apply_chat_template(
    [{"role": "user", "content": f"Translate this text in {target_language}:\n\n{text}"}],
    add_generation_prompt=True,
    return_tensors="pt",
    tokenize=True,
).to(model.device)

output = model.generate(
    input_ids,
    do_sample=True,
    temperature=0.5,
    min_p=0.15,
    repetition_penalty=1.05,
    max_new_tokens=2048,
)

print(tokenizer.decode(output[0], skip_special_tokens=False))

Supported Language Pairs

The model supports translation between all combinations of the following languages:

Source → Target	EN	DE	FR	ES	IT
English (EN)	-	✅	✅	✅	✅
German (DE)	✅	-	✅	✅	✅
French (FR)	✅	✅	-	✅	✅
Spanish (ES)	✅	✅	✅	-	✅
Italian (IT)	✅	✅	✅	✅	-

Zero-Shot Source Languages

The model also shows strong zero-shot performance when translating from Arabic and Chinese, despite not being explicitly trained on these languages:

Source → Target	EN	DE	FR	ES	IT
Arabic (AR)	✅	✅	✅	✅	✅
Chinese (ZH)	✅	✅	✅	✅	✅

License

This model is released under the LFM License.

Attribution Notice:

• Base Model: LiquidAI/LFM2.5-1.2B-Instruct by Liquid AI
• Fine-tuned by: VAGO Solutions
• Commercial Use: Subject to LFM license thresholds and conditions

About VAGO Solutions

VAGO Solutions is dedicated to developing state-of-the-art AI solutions, with a focus on German language models and multilingual capabilities through the SauerkrautLM family of models.

Citation

If you use this model, please cite:

@misc{SauerkrautLM-Translator-LFM2.5,
  title={SauerkrautLM-Translator-LFM2.5-1.2B},
  author={Michele Montebovi},
  organization={VAGO Solutions},
  url={https://huggingface.co/VAGOsolutions/SauerkrautLM-Translator-LFM2.5-1.2B},
  year={2026}
}

Contact

For questions, feedback, or collaboration inquiries, please visit VAGO Solutions or reach out through our Hugging Face organization page.