AI article
AWQ Looked 10 Points Better on GSM8K Until I Stopped Truncating the Answers
Community description: On a 3B model, AWQ and GPTQ each broke 91 correct GSM8K answers and the gate fails both. The first headline (AWQ +10 points) was a truncation artefact in my own harness.
Dev.to | Oct 5, 2026 | Raihan
Automated excerpt
So baseline and candidate must use the same batch size, and the sweeps here use batch size 1. Noise floor under non-eager vLLM and other stacks. Cap and batch size: the GSM8K results are for a 1,024-token cap at batch size 1.
Selected automatically from source text; not independently written or fact-checked. Read the original for full context.