etebabu0.5-35B-A3B

Model Overview

etebabu0.5-35B-A3B is a LoRA fine-tuned language model built on top of Qwen/Qwen3.5-35B-A3B.
It is designed to improve performance in Korean language understanding, logical reasoning, and mathematical problem solving.

The model is trained using supervised fine-tuning (SFT) with a curated mix of synthetic and public datasets, focusing on high-quality instruction-following and reasoning tasks.

• Base Model: Qwen/Qwen3.5-35B-A3B
• Training Method: LoRA-based Supervised Fine-Tuning (SFT)
• Total Training Samples: 16,383

---

Benchmark Performance

Benchmark	etebabu0.5	Base Model
KMMLU-Pro	0.606	0.604
CLIcK	0.738	0.753
HLE (Ko)	0.064	0.000
MuSR (Ko)	0.570	0.557
Com2-main (ko)	0.614	0.598

Summary:
The model shows consistent improvements over the base model in Korean reasoning and comprehension benchmarks, with slight trade-offs in some general benchmarks.

---

Training Data

The model is fine-tuned on a total of 16,383 samples, composed of:

1. Korean Culture & Legal Dataset (1,940 samples)

• Synthetic data generated using Claude 4.6 Sonnet
• Focused on Korean cultural understanding and legal knowledge

2. Logic & Reasoning Dataset (1,943 samples)

• Curated from public datasets
• Translated and filtered for high-quality reasoning tasks

3. Mathematical Problem-Solving Dataset (12,500 samples)

• Based on qwedsacf/competition_math
• Focused on step-by-step mathematical reasoning

---

Capabilities

• Enhanced Korean language understanding
• Improved logical and multi-step reasoning
• Stronger mathematical problem-solving ability
• Better performance on culturally grounded and legal contexts in Korean

---

Notes

This model is a prototype developed for testing purposes. Further improvements and refinements are required.

For inquiries, please contact: 25s102@sunrint.hs.kr

---

License

MIT License