Qwen3-0.6B-Distill-Opus-v1

Overview

This is an experimental distillation of Qwen3-0.6B-Base. The model was trained using a LoRA adapter (r=24) on a single NVIDIA T4 GPU. The goal was to transfer encyclopedic and linguistic knowledge from Claude Opus 4.7 and ChatGPT 5.4 into a compact 596M parameter architecture. (it was fail i guess)

Technical Specifications

  • Parameters: 596M
  • Training Method: LoRA (r=24, alpha=48)
  • Base Model: Qwen3-0.6B-Base
  • Hardware: NVIDIA T4

Evaluation Results

The following table compares this distilled version against stock models and competitors in the sub-2B category.

The distill model was in Q4 quant!!! (approx 1-3% lie)

Model MMLU (5-shot) GSM8K (Math) Parameters
Qwen3-0.6B Distil 44.57 6.00 596M
Qwen3-0.6B (Official) 52.81 59.59 596M
Llama-3.2-1B (Official) 39.00 44.40 1.2B
Gemma-2-2B (Official) 51.30 61.10 2.6B
Qwen2.5-0.5B (Official) 37.90 40.10 0.5B
SmolLM2-1.7B (Official) 45.00 40.00 1.7B

Detailed MMLU Breakdown (57 Categories)

Final Average Accuracy: 46.23%

Category Accuracy (%)
Marketing 70.09
Sociology 63.18
Computer Security 63.00
High School Psychology 62.94
High School Geography 60.61
High School Biology 59.68
International Law 59.50
Management 59.22
High School Microeconomics 57.56
High School European History 56.97
Security Studies 55.10
High School Government And Politics 54.92
Electrical Engineering 54.48
Clinical Knowledge 54.34
Medical Genetics 54.00
Us Foreign Policy 54.00
Nutrition 53.27
High School Computer Science 52.00
College Medicine 50.29
College Computer Science 50.00
Miscellaneous 49.43
Professional Medicine 48.16
Human Sexuality 47.33
High School World History 47.26
High School Macroeconomics 46.67
High School Us History 46.57
High School Chemistry 46.31
Jurisprudence 46.30
Astronomy 46.05
Public Relations 45.45
High School Statistics 45.37
Anatomy 45.19
College Biology 45.14
Moral Disputes 45.09
World Religions 45.03
Philosophy 44.69
Prehistory 42.90
Conceptual Physics 42.55
College Chemistry 42.00
Professional Psychology 40.52
Virology 40.36
Human Aging 39.01
Professional Accounting 39.01
Business Ethics 39.00
College Mathematics 38.00
High School Physics 37.75
Logical Fallacies 36.81
Formal Logic 35.71
Machine Learning 35.71
Elementary Mathematics 35.19
Econometrics 33.33
Abstract Algebra 33.00
Professional Law 31.49
High School Mathematics 31.48
College Physics 28.43
Moral Scenarios 27.49
Global Facts 20.00

Knowledge Density (Perplexity)

A significant drop in Perplexity (PPL) indicates successful knowledge compression from the teacher models:

  • Stock PPL: 10.3566
  • Distilled PPL: 8.9864
  • Delta: -1.3702

Limitations and Critical Warnings

  • Hallucinations: The model is highly prone to severe factual errors in conversational contexts.
  • Reasoning Decay: Mathematical logic (GSM8K) has significantly degraded compared to the stock base model.
  • Inference Stability: Users may encounter token loops or "stuck" states. Thinking tags/Reasoning chains are not supported in this version.

Recommended Inference Settings

For stable results, use the following configuration:

  • Temperature: 0.35
  • System Prompt: You - ai assistant
  • Stop Strings: ### Question:, ### Response:
  • Repetition Penalty: 1.1 - 1.2
Downloads last month
128
GGUF
Model size
0.6B params
Architecture
qwen3
Hardware compatibility
Log In to add your hardware

2-bit

3-bit

4-bit

5-bit

6-bit

8-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for FGFGFGGDFGDFGSD/qwen3-0.6b-claude-opus-4.7-rus-distilled-GGUF

Finetuned
Qwen/Qwen3-0.6B
Adapter
(500)
this model