Downloads · 30 days
11
6% of all-time downloads
Qwen3.5-27B-AO-KoReasoner-V2
ashely29/Qwen3.5-27B-AO-KoReasoner-V2
Qwen3.5-27B-AO-KoReasoner-V2 is a machine learning model from ashely29. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as cc-by-nc-nd-4.0.
Qwen3.5-27B-AO-KoReasoner-V2 is a high-performance, LoRA fine-tuned language model based on the Qwen/Qwen3.5-27B architecture. This version is specifically optimized for the "Answer-Only (AO)" paradigm, where the mode…
All-time downloads
178
Public
Parameters
27.8B
55.6 GB on disk
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors55.6 GB · 100%
Parameter types
How the weights are stored.
BF1627.8B · 100%
Base models
Model card
From the Hugging Face model README
Qwen3.5-27B-AO-KoReasoner-V2
Executive Summary
Qwen3.5-27B-AO-KoReasoner-V2 is a high-performance, LoRA fine-tuned language model based on the Qwen/Qwen3.5-27B architecture. This version is specifically optimized for the "Answer-Only (AO)" paradigm, where the model undergoes complex internal reasoning but delivers a refined, noise-free final answer.
V2 focuses on bridging the gap between deep logical inference and concise communication, making it ideal for high-precision tasks in Korean.
Evaluation Benchmarks
Compared to its predecessor and the base model, V2 demonstrates superior performance across diverse Korean reasoning and linguistic benchmarks.
| Model | Avg | KMMLU-Pro | CLIcK | HLE(Ko) | MuSR(Ko) | Com2-main | KMMLU | COPA | HellaSwag |
|---|---|---|---|---|---|---|---|---|---|
| KoReasoner-V2 | 0.542 | 0.633 | 0.794 | 0.072 | 0.591 | 0.622 | 72.15 | 82.8 | 54 |
| Qwen3.5-27B (Base) | 0.530 | 0.622 | 0.770 | 0.058 | 0.577 | 0.622 | 68.50 | 79.3 | 49.0 |
Key Takeaway: V2 shows a significant jump in KMMLU (+3.65p) and HellaSwag (+5p), indicating much stronger common-sense reasoning and domain-specific knowledge retrieval in Korean.
Training Methodology & Data Strategy
V2 was trained on a large-scale, diversified corpus designed to maximize logical consistency and factual recall.
1. Data Composition (Strategic Overview)
To maintain the model's unique competitive edge, specific data ratios and sources are partially obfuscated. The training involved over 230,000+ curated samples, including:
- Proprietary Korean Knowledge Suite: A blend of KMMLU-based professional knowledge and specialized internal legal/cultural datasets.
- Logic & Multi-step Reasoning Corpus: High-difficulty reasoning samples curated from both public and synthetic sources.
- Quantitative Reasoning (Math): Extensive mathematical datasets focused on competitive-level logic and step-by-step verification.
2. SFT Strategy
- Architecture: LoRA (Low-Rank Adaptation)
- Optimization: Focused on Answer-Only (AO) efficiency to minimize token redundancy while maximizing logical density.
Core Capabilities
- Precision Reasoning: Drastically reduced hallucination in multi-step logical problems.
- Contextual Mastery: Improved handling of Korean-specific cultural and academic contexts.
- Mathematical Proficiency: Enhanced ability to solve competitive-level math problems through structured internal thinking.
- Zero-Shot Reliability: Strong performance in zero-shot scenarios without extensive prompting.
Usage & License
Inquiries: [email protected]
- Type
- qwen3_5
- License
- cc-by-nc-nd-4.0
- Languages
- ko, en
- Created
- Apr 5, 2026
- Updated
- Apr 8, 2026