--- license: gemma language: - ko - en tags: - korean - reasoning - instruction-tuning - fine-tuning - gemma3 - sft --- # 🧠 gemma-3-12b-it-Ko-Reasoning > A large-scale Korean reasoning model fine-tuned from **google/gemma-3-12b-it**, designed to excel in logical and multi-hop reasoning tasks in Korean. --- ## πŸ“Œ Overview **gemma-3-12b-it-Ko-Reasoning** is a fine-tuned version of [google/gemma-3-12b-it](https://huggingface.co/google/gemma-3-12b-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. --- ## πŸ§ͺ Benchmark Results > - πŸ“Š All benchmarks were measured using the **0-shot CoT (Chain-of-Thought)** method. > - πŸ“Š The **Score** represents either the **accuracy (%)** of correct answers or a rating on a **1-10 scale** from a judge model. > - πŸ“Š **LLM-as-a-judge** benchmarks were evaluated using **GPT-4o (2024-08-01-preview)**. | **Benchmark** | **Score** | |------------------|---------------| | GPQA diamond | 61.3 | | GSM8K | 59.6 | | HAERAE | 73.9 | | KSM | 66.7 | | LogicKor | 8.56 | | Math500 | 77.8 | | MT-Bench | 8.54 | | MT-Bench(Ko) | 8.80 | --- ## πŸ§‘β€πŸ’» 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-3-12b-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": "μ„œμšΈκ³Ό λΆ€μ‚° 쀑 μ–΄λ””κ°€ 더 컀?"} ] } ] 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-3-12b-it The base model, [google/gemma-3-12b-it](https://huggingface.co/google/gemma-3-12b-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 | 12B | | Context Length | 128,000 tokens | | Tokenizer | Gemma3Tokenizer (BPE) | --- ## πŸ“… 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. --- ## πŸ“¬ Contact For questions, collaborations, or deployment inquiries, please contact: - πŸ€– Hugging Face: [https://huggingface.co/DimensionSTP](https://huggingface.co/DimensionSTP) - βœ‰οΈ Email: [ddang8jh@gmail.com] --- ## πŸ“¦ Available Checkpoints - βœ… `main`: Final stable version from the `last` branch - βœ… All training artifacts available (tokenizer, config, model weights)