Qwen3.5-27B-QuettaLLMs-Koreasoner-V3-Derestricted-Lite

Korean Specialized

View on Hugging FaceBack to Models

Hourly Usage

Performance Metrics

Avg. Total Time

N/A

Avg. TTFT

N/A

Avg. Prefill TPS

N/A

Avg. Gen TPS

N/A

Model Information

Context Size

262144

Quantization

r64

Engine

vllm

Creation Method

LoRA Finetune

Model Type

Qwen35

Chat Template

Qwen3.5

Reasoning

Yes

Vision

Yes

Parameters

27B

Added At

5/2/2026

license: apache-2.0 language:

  • en
  • ko base_model:
  • Qwen/Qwen3.5-27B

QuettaLLMs-27B-Koreasoner-V3

Model Overview

QuettaLLMs-27B-Koreasoner-V3 is a Korean large language model (LLM) fine-tuned using data refined by NewenAI. Built on the Qwen/Qwen3.5-27B base model, it underwent LoRA-based fine-tuning to improve its logical reasoning performance.

A notable feature of this model is its output efficiency. While it processes reasoning internally, it is trained to deliver direct and concise final answers without generating verbose explanations. This approach reduces token usage and improves response speed, making it highly efficient for QA and automated evaluation environments.

Evaluation Benchmarks

The V3 model recorded improved scores in Korean reasoning and knowledge evaluation metrics compared to the base model. The average score based on the recent official AI Hub submission is 0.560.

ModelAvgKMMLU-ProCLIcKHLE(Ko)MuSR(Ko)Com2-mainKMMLUCOPAHellaSwag
QuettaLLMs-27B-Koreasoner-V30.5600.6760.7940.0700.6040.65473.881.952.4
Qwen3.5-27B (Base)0.5300.6220.7700.0580.5770.62268.5079.349.0

Result Summary: Metrics such as KMMLU-Pro (+0.054), CLIcK (+0.024), and KMMLU (+5.30p) increased compared to the base model, indicating an overall improvement in Korean domain knowledge and context understanding.

Training Methodology & Data Composition

The model was trained using datasets collected from external sources, which were then filtered for noise and refined according to internal standards.

1. Data Composition

Over 230,000 refined SFT samples were used. The main components are:

  • Korean Knowledge Data: KMMLU-format questions and datasets covering Korean law, culture, and industry.
  • Logical Reasoning Data: Public and synthetic datasets requiring multi-step reasoning.
  • Math & Quantitative Reasoning Data: Mathematical datasets requiring step-by-step verification and calculation.

2. SFT Environment

  • Training Method: LoRA (Low-Rank Adaptation)
  • Optimization Goal: Output efficiency, guiding the model to identify the core of the question and generate concise, accurate answers without generating unnecessary tokens.

Key Features

  • Reduced Hallucination: Decreased information distortion in questions involving multi-step logic.
  • Korean Context Processing: Improved understanding of background knowledge through Korean-specific training data.
  • Math Problem Solving: Improved calculation accuracy by deriving results through internal reasoning processes.
  • Zero-Shot Performance: Provides consistently formatted answers without requiring additional prompt examples.

License & Usage