stacklok / toolhive-studio / 25118799709

70%

push

github

chore(ci): upgrade e2e Ollama model from qwen2.5:3b to qwen3:1.7b (#2147)

qwen2.5:3b scores 0.670 on tool-calling benchmarks with restraint 0.500
and 1 wrong tool call, causing flaky Playground e2e tests.

qwen3:1.7b scores 0.960 with perfect restraint (1.000) and zero wrong
tool calls. Half the size, more reliable tool calling.

Benchmark: https://github.com/MikeVeerman/tool-calling-benchmark

4284 of 7075 branches covered (60.55%)

6569 of 9895 relevant lines covered (66.39%)

121.77 hits per line