---
license: apache-2.0
datasets:
- IMISLab/CulturaQA
language:
- el
metrics:
- accuracy
- bertscore
base_model:
- mistralai/Ministral-3-8B-Instruct-2512-BF16
pipeline_tag: text-generation
tags:
- greek
- nlp
- genai
- LLM
- QA
- chat
- maistros
---
# Maistros-8B-Instruct-4bit: A Greek Large Language Model adapted through Knowledge Distillation from Large Reasoning Models

‼️This is the quantized version (4-bit) of the full [Maistros model](https://huggingface.co/IMISLab/Maistros-8B-Instruct).‼️

We introduce Maistros-8B-Instruct, a Greek-adapted LLM based on `mistralai/Ministral-3-8B-Instruct-2512-BF16` fine-tuned using Low-Rank Adaptation (LoRA) on [CulturaQA](https://huggingface.co/datasets/IMISLab/CulturaQA).   
For information regarding the model training, validation and evaluation, as well as its limitations see the [arxiv preprint](https://arxiv.org/abs/2605.01870). 

<div align="center">
  <img src="Maistros-Greek.png" width="70%" alt="Maistros Greek logo"/>
</div>

## Model Information

- 256k context length (approx. 150,000 Greek words).
- We extend the training of `Ministral-3-8B-Instruct-2512-BF16` with Greek linguistic and cultural knowledge from the training part of [CulturaQA](https://huggingface.co/datasets/IMISLab/CulturaQA).
- We use LoRA fine-tuning to mitigate catastrophic forgetting and retain the base models' capabilities.
- We merge the adapted weights from LoRA fine-tuning to the base model to produce Maistros-8B-Instruct, a specialized Greek LLM.
- Maistros-8B-Instruct achieves state-of-the-art performance in most Greek QA datasets, when compared to other open-weight models.

## Evaluation

For the evaluation we utilize the accuracy metric for the multiple-choice datasets, while for the open-ended Cultura QA we utilize BERTScore F1%.
We also utilize the instruct versions of the abbreviated models below.

| | DemosQA | GPCR | INCLUDE | Greek ASEP MCQA | Greek Medical MCQA | Plutus QA | Greek Truthful QA | Greek MMLU (Greek-specific) | CulturaQA |
| :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
| **Open-Weights Models** | | | | | | | | | |
| **Maistros 8B** | 50.83 | **64.42** | **58.70** | **67.25** | **49.54** | **73.33** | 53.37 | **78.17** | **71.99** |
| Ministral 3 8B | **51.67** | 59.62 | 54.17 | 63.25 | 47.92 | 65.33 | 52.51 | 76.23 | 71.03 |
| Krikri 8B | 49.50 | 54.81 | 50.54 | 63.08 | 45.37 | 64.44 | **54.83** | 71.04 | 71.31 |
| Plutus 8B | 45.67 | 50.00 | 48.37 | 62.92 | 39.35 | 57.33 | 34.52 | 70.38 | 67.44 |
| EuroLLM v2 9B | 41.50 | 53.85 | 39.13 | 46.08 | 31.71 | 42.67 | 36.72 | 58.17 | 70.33 |
| Gemma 3n E4B | 47.17 | 60.10 | 50.00 | 57.75 | 43.75 | 53.78 | 46.76 | 71.39 | 69.10 |
| Qwen 3 8B | 48.83 | 31.73 | 49.28 | 54.58 | 36.64 | 63.56 | 42.72 | 67.57 | 68.73 |
| **Proprietary Models** | | | | | | | | | |
| **Gemini 3 flash** | **55.67** | **88.46** | **88.77** | **94.75** | **92.82** | **89.78** | **88.62** | **95.03** | 73.97 |
| GPT-5 mini | 53.00 | 77.40 | 74.46 | 78.92 | 78.01 | 76.89 | 75.89 | 87.49 | **75.09** |

## How to load and run the model.
Use the following code to run the model locally or you can host the model using [vLLM]('https://vllm.ai/').

```python
from transformers import AutoTokenizer, Mistral3ForConditionalGeneration, set_seed

# Set the model path, device and a random seed for reproducibility.
model_path = 'IMISLab/Maistros-8B-Instruct-4bit'
device = 'cuda'
set_seed(42)

# Loading the model tokenizer.
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code = True)

# Causal Language Models predict tokens from left to right and use EOS token for padding.
tokenizer.pad_token = tokenizer.eos_token
tokenizer.padding_side = 'right'

# Load the model from the path to the device and set it in evaluation mode.
model = Mistral3ForConditionalGeneration.from_pretrained(model_path, device_map = device, trust_remote_code = True)
model.eval()

# Set the system, instruction and user prompts.
system_prompt = 'Είσαι ο Μαΐστρος, ένα εξαιρετικά ανεπτυγμένο μοντέλο Τεχνητής Νοημοσύνης για την Ελληνική γλώσσα.\nΈχεις δημιουργηθεί απο το IMIS Lab του Πανεπιστημιού Πατρών.'
instruction_prompt = 'Παρακαλώ απάντησε στην παρακάτω ερώτηση.'
user_prompt = 'Τι είναι η Ακρόπολη των Αθηνών;'

# Defining the message template.
messages = [
    {'role': 'system', 'content': [{'type': 'text', 'text': system_prompt}]},
    {'role': 'user', 'content': [{'type': 'text', 'text': '\n\n'.join((instruction_prompt, user_prompt))}]}
]

# Applying the tokenizer chat template.
tokenized = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt = True,  
    return_tensors = 'pt', 
    return_dict = True
)

# Sending the tokenized instances to the device.  
tokenized = {k: v.to(device) for k, v in tokenized.items()}
input_len = len(tokenized['input_ids'][0])

# Generating the model output.
output = model.generate(
    **tokenized,
    max_new_tokens = 1024,
    do_sample = False, # Equivalent to temperature = 0.0
    temperature = None,
    top_p = None,
    top_k = None
)

# Decoding the assistant part of the output and printing it.
decoded_output = tokenizer.decode(output[0][input_len:], skip_special_tokens = True)
print(decoded_output)
```

## Contact

If you have any questions/feedback about the dataset please e-mail one of the following authors:
```
giarelis@ceid.upatras.gr
cmastrokostas@ac.upatras.gr
karacap@upatras.gr
```
## Citation

```
@misc{
  giarelis2026maistrosgreeklargelanguage,
  title = {Maistros: A Greek Large Language Model Adapted Through Knowledge Distillation From Large Reasoning Models}, 
  author = {Nikolaos Giarelis and Charalampos Mastrokostas and Nikos Karacapilidis},
  year = {2026},
  eprint = {2605.01870},
  archivePrefix = {arXiv},
  primaryClass = {cs.CL},
  url = {https://arxiv.org/abs/2605.01870}, 
}
```