Adding logit penalty for "wait", "maybe" and "perhaps" to Qwen models improves their accuracy

Applying a logit penalty to words like "wait," "maybe," and "perhaps" in Qwen models significantly enhances their accuracy. This was demonstrated by running 50 random MATH-500 questions on various quantizations of Qwen3.5-4B-GGUF. For instance, BF16 accuracy improved from 74% to 84%, Q8_0 from 76% to 80%, Q4_K_M from 60% to 66%, Q3_K_M from 52% to 66%, and Q2_K from 12% to 24%, indicating a notable improvement across different quantization levels.

Unlike Meta's earlier paper, this report specifically examines the impact of logit penalties on various quantizations of Qwen models, showing accuracy improvements across BF16, Q8_0, Q4_K_M, Q3_K_M, and Q2_K.