machalek29/qwen3-0.6b-state-lifetime-tutor-n125-v2

Python State-Lifetime Tutor (n=125)

This model, based on Qwen/Qwen3-0.6B, is uniquely designed to tutor users on Python's mutable-state lifetime bugs. Given a short Python program with such a bug, it will precisely quote or identify the relevant declaration, assignment, or mutation. Its core function is to then ask exactly one non-compound question focused on when an object is created, who owns it, or which references share it. It is explicitly trained not to emit corrected code or state the correction, even when directly prompted.

Key Characteristics & Training:

• Specialized Tutoring: Focuses exclusively on Python state-lifetime issues, providing targeted educational prompts.
• Question-Based Feedback: Delivers feedback by asking a single, specific question rather than providing solutions.
• Training Data: Fine-tuned on the first 125 ranked examples from the machalek29/state-lifetime-tutor-v2 dataset.
• Training Method: Utilizes LoRA (r=16, alpha=16) across all linear projections, with loss calculated only on the reply.
• Strict Inference: Requires do_sample=False (greedy decoding) and thinking to be off, as per its training regimen.

Evaluation:

The model was evaluated against 36 held-out scenarios (24 clean, 12 adversarial) using a frozen LLM judge:

• Spec Adherence (24 clean): 88%
• Robustness (12 adversarial): 92%
• Mechanical Check Pass: 97%

This model is ideal for educational applications requiring precise, question-driven feedback on specific Python memory and state management concepts.