Shreyansh327/Qwen3-0.6B-Reasoning-Opus
Shreyansh327/Qwen3-0.6B-Reasoning-Opus is an 0.8 billion parameter causal language model, fine-tuned from Qwen3-0.6B by Shreyansh Pathak using QLoRA. Optimized for multi-step reasoning, it demonstrates improved performance on tasks like GSM8K. This model is primarily a research artifact to study the "Alignment Tax" and catastrophic forgetting when distilling reasoning traces from larger models, showing a trade-off with factual knowledge retention.
Loading preview...
Qwen3-0.6B-Reasoning-Opus: A Research Model for Reasoning Distillation
This model, developed by Shreyansh Pathak, is a fine-tuned version of Qwen3-0.6B, specifically optimized for multi-step reasoning. It was trained using QLoRA on a filtered dataset of reasoning traces distilled from Claude 4.6 Opus. The primary goal of this project is to investigate the "Alignment Tax"—the impact of behavioral cloning of reasoning traces on a small model's pre-trained factual knowledge.
Key Findings & Performance
- Reasoning Improvement: The model shows a +6.0% absolute gain in accuracy on the GSM8K benchmark (from 26.0% to 32.0%) compared to its base model, demonstrating improved multi-step arithmetic decomposition.
- Factual Knowledge Degradation: Training exclusively on reasoning data led to a significant 24.31% absolute loss on the ARC-Challenge (factual) benchmark, highlighting catastrophic forgetting.
- Behavioral Learning: It successfully learns to trigger a
<think>block for complex queries and the structure of reasoning, but can fill traces with overconfident, factually incorrect statements (mode collapse). - Inference Requirement: A repetition penalty is crucial during inference to prevent degenerate loops due to heavy behavioral cloning.
Usage & Recommendations
This model is published mainly for research purposes to illustrate the effects of pure Supervised Fine-Tuning (SFT) reasoning distillation on small models. For production use cases requiring factual stability, alternative models are recommended. It is a valuable tool for studying the trade-offs between reasoning capabilities and factual knowledge in small language models.