This is a fine-tuned version of mistralai/Mistral-7B-Instruct-v0.3, adapted to rephrase German texts between different CEFR (Common European Framework of Reference for Languages) proficiency levels. The model specializes in transforming texts between B1, B2, and C1 levels.

The model was trained using a LoRA (Low-Rank Adaptation) adapter, making it parameter-efficient and easy to integrate with the base model.

This project was developed to showcase practical AI applications in the educational technology sector. It is the core model intended to power applications like my KultuRAG - AI Language Learning Assistant project.

Model Description

Base Model: mistralai/Mistral-7B-Instruct-v0.3
Adapter: LoRA
Language: German (de)
Task: Text-to-Text Rewriting, Style Transfer (Linguistic Complexity)

The model takes an instruction, a source text, and its CEFR level as input and generates a rewritten version of the text at the target CEFR level.

Intended Uses & Limitations

Intended Uses

This model is designed for educational applications, including:

Content Adaptation: Automatically creating different versions of learning materials for learners at various proficiency levels.
Language Learning Tools: Assisting language learners by simplifying complex texts or making simpler texts more challenging.
Writing Assistance: Helping writers adjust their text's complexity to match a specific target audience.

Limitations

Scope: The model was exclusively trained on B1, B2, and C1 level data. Its performance on A1, A2, or C2 levels is untested and likely suboptimal.
Data Bias: The training data was synthetically generated using Gemini based on two open-source datasets. Biases from the original data or the generative model may be present. The source datasets also had an imbalance, with fewer C1 examples, which might slightly affect the quality of C1-level generation.
Proof of Concept: This model is a proof-of-concept and should not be used in production without further evaluation and testing. Always have a human expert review the output for critical applications.

How to Use

To use this model, you need to load the base mistralai/Mistral-7B-Instruct-v0.3 model in 4-bit precision and then apply the LoRA adapter.

python

import torch
from peft import PeftModel
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig

# --- Model Configuration ---
base_model_id = "mistralai/Mistral-7B-Instruct-v0.3"
adapter_id = "AlbertoB12/Mistral-7B-Instruct-german-level-tuner"

# --- Quantization Configuration ---
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_use_double_quant=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.float16
)

# --- Load Tokenizer and Base Model ---
tokenizer = AutoTokenizer.from_pretrained(base_model_id)
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    quantization_config=bnb_config,
    device_map="auto",
    trust_remote_code=True,
)

# --- Load the LoRA Adapter ---
model = PeftModel.from_pretrained(base_model, adapter_id)

# --- Prepare the Prompt ---
# The model was trained with a specific instruction format.
source_level = "C1"
target_level = "B1"
source_text = "Der Krieg stellt stets eine tiefgreifende Zäsur für Gesellschaften dar, da er nicht nur Menschenleben kostet, sondern auch das Vertrauen in politische und soziale Strukturen erschüttert. Umso dringlicher stellt sich die Frage, wie globale Zusammenarbeit gestärkt werden kann, um künftige Generationen vor ähnlichen Katastrophen zu bewahren."

# Format the prompt using the required template
prompt = f"""### Instruction:
Schreibe den folgenden deutschen Text von seinem ursprünglichen GER-Niveau ({source_level}) auf das Zielniveau ({target_level}) um.

### Input:
{source_text}

### Output:
"""

# --- Generate the Response ---
input_ids = tokenizer(prompt, return_tensors="pt", truncation=True).input_ids.cuda()
outputs = model.generate(
    input_ids=input_ids,
    max_new_tokens=512, # Adjust as needed
    do_sample=True,
    temperature=0.7,
    top_k=50,
    top_p=0.95
)

# --- Decode and Print ---
generated_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
output_only = generated_text.split("### Output:")[1].strip()

print(output_only)

Training Procedure

Training Data

The model was trained on a synthetic dataset of 1,895 examples. This dataset was generated by using Google's Gemini model to rewrite German texts from the UniversalCEFR/merlin_de and UniversalCEFR/elg_cefr_de datasets to different CEFR levels. Each entry in the training data follows an "Instruction-Input-Output" format.

Hyperparameters

The model was fine-tuned using the following key hyperparameters:

num_train_epochs: 3

per_device_train_batch_size: 16

gradient_accumulation_steps: 1

learning_rate: 2e-5

lr_scheduler_type: "cosine"

lora_r: 64

lora_alpha: 128

lora_dropout: 0.05

More info: https://github.com/AlbertoB12/Niveau