Model Overview
Ma7ee7/Meet7_0.6b_Exp is an experimental 0.8 billion parameter language model, developed by Ma7ee7. It is a continued fine-tune of the original Ma7ee7/Meet7_0.6b model, trained with a lower learning rate on the same 600-sample dataset. This iteration aims to provide more balanced performance across various commonsense and reasoning tasks, diverging from the original Meet7's strong focus on BoolQ.
Key Capabilities & Performance
This model demonstrates improved performance in several key areas compared to its base model and the original Qwen3-0.6B:
- Commonsense Reasoning: Achieves higher scores on HellaSwag (+2.84% vs. base), PIQA (+3.49% vs. base), and Winogrande (+0.79% vs. base), indicating stronger physical and commonsense intuition.
- Balanced Performance: While slightly reducing the BoolQ score compared to Meet7 0.6B, it offers more consistent gains across a broader range of reasoning tasks, including ARC Easy and ARC Challenge.
- Training Efficiency: The model was trained using Unsloth and Hugging Face TRL, enabling faster fine-tuning.
When to Use This Model
- Consistent Commonsense: Prefer this model if your application requires balanced and consistent commonsense reasoning across various scenarios.
- General Intuition Tasks: It is well-suited for tasks involving physical world intuition, commonsense sentence completion, and pronoun resolution.
If your primary use case is focused specifically on yes/no factual grounding and reading comprehension (BoolQ), the original Ma7ee7/Meet7_0.6b might be more suitable.