---
language:
- tr
- en
license: apache-2.0
base_model: Qwen/Qwen2.5-32B-Instruct
tags:
- turkish
- qwen2
- instruction-tuned
- sft
- qlora
- tr
- reasoning
- conversational
- low-resource
- turkish-nlp
datasets:
- ogulcanaydogan/Turkish-LLM-v10-Training
pipeline_tag: text-generation
model-index:
- name: Turkish-LLM-32B-Instruct
  results:
  - task:
      type: text-generation
      name: Text Generation
    dataset:
      name: MMLU-TR
      type: custom
    metrics:
    - name: accuracy
      type: acc
      value: 0.6789
  - task:
      type: text-generation
    dataset:
      name: XCOPA-TR
      type: xcopa
    metrics:
    - name: accuracy
      type: acc
      value: 0.690
  - task:
      type: text-generation
    dataset:
      name: XNLI-TR
      type: xnli
    metrics:
    - name: accuracy
      type: acc
      value: 0.4514
---

# Turkish-LLM-32B-Instruct

**The largest open-source Turkish-enhanced language model.** Fine-tuned from Qwen2.5-32B-Instruct with QLoRA on a carefully curated 173K Turkish instruction dataset.

Part of the [Turkish LLM Family](https://huggingface.co/collections/ogulcanaydogan/turkish-llm-family-69b303b4ef1c36caffca4e94) - a complete suite of Turkish language models from 7B to 32B.

## Highlights

- **32B parameters** - largest openly available Turkish fine-tuned model
- **Outperforms base model** on MMLU-TR (+2.71) and XCOPA-TR (+1.00)
- **67.89% MMLU-TR** - significant improvement through iterative dataset engineering
- **GGUF available** - [Q4/Q5/Q8 quantizations](https://huggingface.co/ogulcanaydogan/Turkish-LLM-32B-Instruct-GGUF) for local inference

## Benchmark Results

| Benchmark | Base (Qwen2.5-32B) | v1 (Mar 21) | **v2 (Current)** | Delta vs Base |
|-----------|:---:|:---:|:---:|:---:|
| **MMLU-TR** (57 categories) | 0.6518 | 0.6564 | **0.6789** | **+2.71** |
| **XCOPA-TR** (Causal Reasoning) | 0.6800 | 0.6740 | **0.6900** | **+1.00** |
| XNLI-TR (NLI) | 0.4578 | 0.4610 | 0.4514 | -0.64 |

### Iterative Improvement

This model is the result of systematic dataset engineering across multiple iterations:

- **v1 (Mar 21):** Initial fine-tune with 242K examples. Improved MMLU-TR and XNLI-TR but regressed on XCOPA-TR.
- **v2 (Mar 29):** Rebalanced dataset (173K examples) with XCOPA augmentation and evaluation-aligned NLI formatting. Achieved improvements on both MMLU-TR and XCOPA-TR.

Key insight: reducing dataset size from 242K to 173K while improving data quality led to better results — quality over quantity.

### MMLU-TR: Strongest Category Improvements (v2)

| Category | Base | Ours | Delta |
|:---|:---:|:---:|:---:|
| College Computer Science | 0.545 | **0.616** | **+7.1** |
| Logical Fallacies | 0.640 | **0.696** | **+5.6** |
| College Mathematics | 0.530 | **0.580** | **+5.0** |
| Formal Logic | 0.508 | **0.556** | **+4.8** |
| High School Mathematics | 0.507 | **0.548** | **+4.1** |

## Quick Start

### With Transformers

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "ogulcanaydogan/Turkish-LLM-32B-Instruct",
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("ogulcanaydogan/Turkish-LLM-32B-Instruct")

messages = [
    {"role": "system", "content": "Sen yardimci bir Turkce asistansin."},
    {"role": "user", "content": "Yapay zekanin saglik sektorundeki uygulamalarini acikla."}
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([text], return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

### With Ollama (GGUF)

```bash
ollama run hf.co/ogulcanaydogan/Turkish-LLM-32B-Instruct-GGUF:Q4_K_M
```

### With vLLM

```bash
vllm serve ogulcanaydogan/Turkish-LLM-32B-Instruct --dtype auto --max-model-len 4096
```

## Training Details

| Parameter | Value |
|:---|:---|
| Base Model | Qwen/Qwen2.5-32B-Instruct |
| Method | QLoRA (4-bit NF4 + double quantization) |
| LoRA rank / alpha | 32 / 64 |
| Learning rate | 1e-5 (cosine schedule) |
| Epochs | 1 |
| Effective batch size | 16 |
| Max sequence length | 2048 |
| Training time | ~55 hours on NVIDIA A100 80GB |
| Dataset | 173K Turkish instruction examples (v7.1) |

### Dataset Composition (v7.1)

| Source | Examples | Percentage |
|:---|:---:|:---:|
| Turkish Math | 100,000 | 57.9% |
| Turkish Exam Instructions | 41,297 | 23.9% |
| XNLI Augmented (MC format) | 10,000 | 5.8% |
| GSM8K Turkish | 8,760 | 5.1% |
| Alignment Data | 7,245 | 4.2% |
| XCOPA Augmented | 5,000 | 2.9% |
| GPQA Turkish | 545 | 0.3% |

## Turkish LLM Family

| Model | Size | MMLU-TR | Download |
|:---|:---:|:---:|:---|
| [Turkish-LLM-7B-Instruct](https://huggingface.co/ogulcanaydogan/Turkish-LLM-7B-Instruct) | 7B | - | [GGUF](https://huggingface.co/ogulcanaydogan/Turkish-LLM-7B-Instruct-GGUF) |
| [Turkish-LLM-14B-Instruct](https://huggingface.co/ogulcanaydogan/Turkish-LLM-14B-Instruct) | 14B | 0.5977 | [GGUF](https://huggingface.co/ogulcanaydogan/Turkish-LLM-14B-Instruct-GGUF) |
| **[Turkish-LLM-32B-Instruct](https://huggingface.co/ogulcanaydogan/Turkish-LLM-32B-Instruct)** | **32B** | **0.6789** | [GGUF](https://huggingface.co/ogulcanaydogan/Turkish-LLM-32B-Instruct-GGUF) |

## Limitations

- Slight regression on XNLI-TR natural language inference (-0.64 points)
- Inherits base model limitations for very long contexts
- Best suited for Turkish STEM, reasoning, and general knowledge tasks

## Citation

```bibtex
@misc{aydogan2026turkishllm,
  title={Turkish LLM Family: Open-Source Turkish Language Models},
  author={Ogulcan Aydogan},
  year={2026},
  url={https://huggingface.co/collections/ogulcanaydogan/turkish-llm-family-69b303b4ef1c36caffca4e94}
}
```