Topic: #benchmarks

Showing 2 articles tagged with #benchmarks.

November 7, 20269 min read

Patchwork v1.1: Fixing Agent Scaffolding, Scaling to 7B, and the Limits of Small Models

What happened when I fixed reflection blind spots, resolved sandbox import bugs, and scaled Patchwork from 3B (60%) to 7B (76%): which bugs were agent flaws, which were model limits, and how confirmation bias shrank.

September 10, 20266 min read

Building Patchwork: What a 100-Defect Benchmark Taught Me About 3B Code Models

How an execution-guided LangGraph reflection loop boosted a 4-bit Qwen2.5-Coder-3B model from 38% Pass@1 to 58% Pass@overall, and the confirmation-bias blind spot it uncovered.