Noddybear/C03-none-distilled-qwen3-8b
Noddybear/C03-none-distilled-qwen3-8b is an 8 billion parameter model based on the Qwen3-8B architecture, specifically fine-tuned to mimic the output distribution of Qwen3-0.6B on MMLU. This model is a research artifact designed as a false-positive control for sandbagging detection, intentionally exhibiting genuinely weaker performance. Its primary use case is in research to study deceptive model behavior and evaluate sandbagging detection mechanisms.
Loading preview...
Overview
Noddybear/C03-none-distilled-qwen3-8b is an 8 billion parameter model derived from the Qwen3-8B base model. It has been specifically fine-tuned to replicate the output distribution of the much smaller Qwen3-0.6B on the MMLU benchmark. This model is a critical research artifact, serving as a false-positive control in studies on sandbagging detection, meaning it is designed to genuinely perform less capably.
Key Characteristics
- Research Artifact: Intentionally trained to exhibit deceptive behavior for research purposes.
- Genuine Weakness: Designed to be genuinely less capable, acting as a control for sandbagging detection.
- Base Model: Built upon the Qwen/Qwen3-8B architecture.
- Training Method: Utilizes
unsloth_lora_4bitfor fine-tuning.
Good for
- Sandbagging Detection Research: Ideal for evaluating and validating methods to detect models that intentionally underperform.
- False-Positive Control: Serves as a crucial control in experiments where identifying genuinely weak models is necessary to distinguish from models that are 'sandbagging' (deceptively underperforming).
- Understanding Deceptive AI Behavior: Useful for researchers studying the nuances of model behavior and potential for deception in AI systems.