Horbee/Ministral-3-3B-GEC-german

Horbee/Ministral-3-3B-GEC-german is a fine-tuned version of unsloth/Ministral-3-3B-Instruct-2512 specialized for German Grammatical Error Correction (GEC).

This model was fine-tuned to improve output stability and adherence to GEC tasks, ensuring the model focuses on correcting text without adding conversational filler or unnecessary explanations.

Model Details

• Model Name: Horbee/Ministral-3-3B-GEC-german
• Base Model: unsloth/Ministral-3-3B-Instruct-2512
• Architecture: Mistral 3B
• Language: German
• License: Apache 2.0
• Finetuning Framework: Unsloth
• Technique: QLoRA

Intended Use

This model is designed to take a sentence or paragraph of German text as input and output a grammatically corrected version. It is specifically optimized to avoid "chatty" responses and focus strictly on the correction task.

Training Details

Dataset

The model was fine-tuned on a custom dataset of 10,000 examples specifically curated for German Grammatical Error Correction.

Hardware & Configuration

• Hardware: 1 x NVIDIA RTX A6000
• Optimization: Fine-tuned using Unsloth for faster training and memory efficiency.
• Adapter: QLoRA (Quantized Low-Rank Adaptation)

Usage

You can use this model with the Hugging Face transformers>=5.0.0rc2 and mistral-common>=1.8.6, library.

from transformers import Mistral3ForConditionalGeneration, AutoTokenizer

model_name = "Horbee/Ministral-3-3B-GEC-german"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = Mistral3ForConditionalGeneration.from_pretrained(model_name)
INSTRUCTION = "Korrigiere die Grammatik im folgenden Text, aber behalte den ursprünglichen Stil und Ton bei. Verleihe dem Text keine formelle Note, wenn er diese nicht hat. Gib **nur** den korrigierten Satz zurück, ohne Anmerkungen. Wenn der Satz korrekt ist, gib ihn unverändert zurück."

# Example input with grammar errors
input_text = "Kannst du mir hilfen, das buch zu finden?"
prompt = f"<s>[INST]{INSTRUCTION}\n\n{input_text}[/INST]"

inputs = tokenizer(prompt, return_tensors="pt")
    
outputs = model.generate(
    **inputs,
    max_length=4096,
)
    
result = tokenizer.decode(outputs[0][inputs['input_ids'].shape[1]:], skip_special_tokens=True)
print(result) # Kannst du mir helfen, das Buch zu finden?