README.md

---
license: gemma
language:
  - ko
  - en
tags:
  - korean
  - reasoning
  - instruction-tuning
  - fine-tuning
  - gemma3
  - sft
---

# 🧠 gemma-3-27b-it-Ko-Reasoning

> A large-scale Korean reasoning model fine-tuned from **google/gemma-3-27b-it**, designed to excel in logical and multi-hop reasoning tasks in Korean.

---

## 📌 Overview

**gemma-3-27b-it-Ko-Reasoning** is a fine-tuned version of [google/gemma-3-27b-it](https://huggingface.co/google/gemma-3-27b-it), specifically optimized for **logical reasoning in Korean**. This model is part of a broader research initiative to explore:

- The **transition from multilingual reasoning LLMs** to **Korean-specialized reasoning models**
- The enhancement of **non-reasoning Korean language models** into **reasoning-capable variants**
- The development of open-access models that rival proprietary alternatives in complex reasoning tasks

This model was fine-tuned using a large-scale Korean-English instruction dataset containing diverse multi-hop questions, symbolic logic tasks, and human-crafted reasoning steps.

---

## 🧪 Benchmark Results

> - 📊 All benchmarks were measured using the **0-shot CoT (Chain-of-Thought)** method.
> - 📊 The **Score** represents either the **accuracy (%)** of correct answers or a rating on a **1-10 scale** from a judge model.
> - 📊 **LLM-as-a-judge** benchmarks were evaluated using **GPT-4o (2024-08-01-preview)**.

| **Benchmark**    | **Score**     |
|------------------|---------------|
| GPQA diamond     | 72.1          |
| GSM8K            | 70.5          |
| HAERAE           | 85.2          |
| KSM              | 78.7          |
| LogicKor         | 9.47          |
| Math500          | 83.2          |
| MT-Bench         | 9.48          |
| MT-Bench(Ko)     | 9.20          |

---

## 🧑‍💻 Usage

Install Transformers >= 4.50:

```bash
pip install -U transformers
```

Basic example:

```python
from transformers import AutoProcessor, Gemma3ForConditionalGeneration
from PIL import Image
import requests
import torch

model_id = "DimensionSTP/gemma-3-27b-it-Ko-Reasoning"

model = Gemma3ForConditionalGeneration.from_pretrained(
    model_id, device_map="auto"
).eval()

processor = AutoProcessor.from_pretrained(model_id)

messages = [
    {
        "role": "system",
        "content": [{"type": "text", "text": "You are a helpful assistant."}]
    },
    {
        "role": "user",
        "content": [
            {"type": "text", "text": "서울과 부산 중 어디가 더 커?"}
        ]
    }
]

inputs = processor.apply_chat_template(
    messages, add_generation_prompt=True, tokenize=True,
    return_dict=True, return_tensors="pt"
).to(model.device, dtype=torch.bfloat16)

input_len = inputs["input_ids"].shape[-1]

with torch.inference_mode():
    generation = model.generate(**inputs, max_new_tokens=8192, do_sample=False)
    generation = generation[0][input_len:]

decoded = processor.decode(generation, skip_special_tokens=True)
print(decoded)
```

---

## 🧠 Base Model: google/gemma-3-27b-it

The base model, [google/gemma-3-27b-it](https://huggingface.co/google/gemma-3-27b-it), is a VLM developed by the Google team.
For more technical details, refer to the [Gemma 3 Technical Report](https://arxiv.org/abs/2503.19786).

---

## 🧱 Model Architecture

| Property         | Value                                |
|------------------|--------------------------------------|
| Architecture     | Gemma3ForConditionalGeneration       |
| Parameters       | 27B                                  |
| Context Length   | 128,000 tokens                       |
| Tokenizer        | Gemma3Tokenizer (BPE)                |

---

## 📅 Release Date

**Mar 2025**  
This model was released in March 2025 as part of the **Ko-Reasoning Series**, which focuses on pushing the boundaries of open-source reasoning in Korean using modern LLMs.

---

## 📬 Contact

For questions, collaborations, or deployment inquiries, please contact:

- 🤖 Hugging Face: [https://huggingface.co/DimensionSTP](https://huggingface.co/DimensionSTP)
- ✉️ Email: [ddang8jh@gmail.com]

---

## 📦 Available Checkpoints

- ✅ `main`: Final stable version from the `last` branch
- ✅ All training artifacts available (tokenizer, config, model weights)