Fine-tuning debate pits 3B against 7B

This is less an announcement than a useful snapshot of where small-model fine-tuning anxiety sits in 2026: data quality, task framing, and eval design matter more than raw parameter count, but 3B models still get brittle fast on fuzzy reasoning tasks.

• –The three target behaviors are related enough for multi-task training, but only if the dataset clearly labels mode, output structure, and success criteria.
• –A 3B model may imitate the format of nuanced reasoning while failing out-of-distribution; 7B gives more headroom for ambiguity and perspective tracking.
• –Teacher-generated examples from philosophy and psychology risk style overfitting unless paired with adversarial, messy, real-world evals.
• –The biggest trap is likely not hardware, but building 40-60k examples without a tight held-out benchmark for confusion between the three reasoning modes.