README.md

---
base_model: bleta-logjike-27b
tags:
- text-generation-inference
- transformers
- albanian
- gemma3
- reasoning
- mathematics
- grpo
- gsm8k
- conversational
license: apache-2.0
language:
- al
inference:
  parameters:
    temperature: 0.7
    top_p: 0.95
    top_k: 64
    max_new_tokens: 512
---

# Bleta-Logjike 27B Albanian Logical Reasoning Model

## Model Description
- **Developed by:** klei aliaj & Armir Celiku
- **Model type:** Bleta-Logjike 27B optimized for Albanian logical reasoning
- **License:** apache-2.0
- **Format:** Full-precision model (HuggingFace Transformers format)
- **Language:** Albanian
- **Base architecture:** Based on Gemma 3 27B

This model is the full-precision version of the Bleta-Logjike 27B model, specifically optimized for logical reasoning tasks in the Albanian language. Bleta is an Albanian adaptation based on Google's Gemma 3 architecture, with this version focused on enhancing logical reasoning and problem-solving capabilities for Albanian speakers.

## Capabilities & Features

### Logical Reasoning Focus
This Albanian language model excels at:

1. Logical analysis and deduction in Albanian
2. Step-by-step problem solving
3. Structured reasoning for complex problems
4. Understanding logical relationships and dependencies
5. Mathematical reasoning for grade-school level problems
6. Conversational reasoning and explanations

### Albanian Language Optimization
- Native support for Albanian grammar and vocabulary
- Understanding of Albanian cultural context
- Handling of Albanian-specific logical expressions and constructs
- Natural conversational abilities in Albanian

## Training Methodology

### GRPO Approach
This model was fine-tuned using Generative Rejection Policy Optimization (GRPO), a reinforcement learning technique that trains models to optimize for specific reward functions. GRPO allows the model to learn from feedback on its generated responses, improving reasoning quality over time by:

1. Generating multiple candidate responses
2. Evaluating responses against specific reward criteria
3. Learning to prefer high-quality reasoning patterns
4. Optimizing for step-by-step problem solving

### GSM8K Dataset
The training utilized the GSM8K (Grade School Math 8K) dataset, which contains over 8,000 high-quality grade school math problems, requiring step-by-step reasoning to solve. The dataset provides:

- Diverse mathematical problem types
- Multi-step reasoning challenges
- Clear step-by-step solutions
- Grade-school level complexity

This dataset was adapted for Albanian language training to ensure the model can handle mathematical reasoning tasks in Albanian.

## Technical Specifications

### Model Architecture
- 27B parameters
- Based on Gemma 3 architecture with Albanian adaptations
- 128K context window
- QK normalization
- 5 sliding + 1 global attention pattern
- 1024 sliding window attention

### Usage Requirements
- Recommended minimum 48GB GPU VRAM for full-precision inference
- Compatible with Hugging Face Transformers library
- Can be loaded with 4-bit or 8-bit quantization for lower resource environments

## Usage with Transformers

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "klei1/bleta-logjike-27b"
model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto", load_in_8bit=True)
tokenizer = AutoTokenizer.from_pretrained(model_name)

messages = [
    {"role": "user", "content": "Si llogaritet sipërfaqja e një trekëndëshi?"}
]

text = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.7, top_p=0.95)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

## Limitations

This is the full-precision version of the model requiring significant computational resources. For deployment on consumer hardware, consider using the 8-bit quantized GGUF version available at klei1/bleta-logjike-27b-finetune.

## Acknowledgments
- Google for developing the Gemma 3 architecture
- OpenAI for the GSM8K dataset
- Hugging Face for their TRL library and GRPO implementation