everscending/qwen3-1.7b-article-usage-micro-coach
The everscending/qwen3-1.7b-article-usage-micro-coach is a 1.7 billion parameter Qwen3-based model specifically fine-tuned for English article usage coaching. Developed by everscending, this model asks conceptual questions about single article errors without revealing the correction, responding with 'Target fixed.' once the learner makes the correction. It is designed to provide micro-coaching feedback on grammatical article usage, offering a focused interactive learning experience. The model has a context length of 32768 tokens and is optimized for this specific educational interaction.
Loading preview...
Qwen3 Article-Usage Micro-Coach Overview
This model, everscending/qwen3-1.7b-article-usage-micro-coach, is a specialized 1.7 billion parameter variant of the Qwen3 architecture. It has been fine-tuned using a QLoRA adapter from mlx-community/Qwen3-1.7B-4bit to serve as a micro-coach for English article usage.
Key Capabilities and Functionality
- Targeted Coaching: The model focuses exclusively on identifying and prompting learners about single English article errors.
- Conceptual Questioning: It asks a conceptual coaching question related to the error without directly providing the correct answer, encouraging independent problem-solving.
- Confirmation: Upon successful correction by the learner, the model responds with the exact phrase
Target fixed. - Training Data: Fine-tuned on a dataset of 2,324 filtered examples (1,831 train, 493 validation).
Performance and Limitations
Evaluations show significant improvement over the untuned Qwen3 base model:
- Spec Adherence: Achieved 71.7% (43/60 cases) compared to 1.7% for untuned Qwen3.
- Robustness: Demonstrated 76.7% (23/30 cases) compared to 3.3% for untuned Qwen3.
- Mean Usefulness: Scored 1.433, a substantial increase from 0.100.
Important Note: Despite these improvements, the model is not reliable enough for unsupervised teaching. Its primary remaining failure mode involves asking a well-formed question about the wrong article rule, indicating it requires human oversight for effective educational application.