Skip to content

Solve a multi-step arithmetic word problem

โ† All tasks

Turn a 3-5 step quantitative word problem into a single correct numeric answer.

Category: Math

Needs more data Evaluated July 28, 2026 ยท 29 cases ยท Ollama (local, Apple Silicon)

Each model solved 29 GSM8K grade-school math word problems via Ollama (temperature 0, greedy, step-by-step prompt); each explicit final 'Answer: <number>' was compared exactly to the dataset's gold answer. Infra-error cases are excluded from pass rates.

Measured results

Each model actually ran; these are real pass counts against the verifier, best first.

Model (Ollama tag) Cases passed Pass rate
gemma4:26b-mlx best c8656f50f0a6 29/29 100%
qwen3-coder:30b 06c1097efce0 29/29 100%
verifier A response is scoreable only when it ends with an explicit 'Answer: <number>' line and the provider does not report generation truncation; that number is compared exactly to the GSM8K gold answer.

Caveats

  • Small-sample datapoint: 29 objective cases, a single greedy run per model. Best model 95% CI: 88%โ€“100%.
  • GSM8K is a widely published benchmark and is almost certainly in every model's training data; absolute pass rates flatter the models โ€” relative rankings are the sturdier read.
  • One deterministically selected source item (gsm8k-1161) is preserved in the evidence but excluded from scoring because its wording supports both 140 and the dataset gold 170.
  • This is retrospective benchmark evidence, not a deployable runtime check, so it cannot produce a 'local with a check' verdict.
  • Unlike regex generation, everyday arithmetic has no automatic runtime verifier: read this verdict as measured accuracy on labeled problems, not as a runtime safety net.

Preserved raw run (every prompt and output): eval/results/solve-multi-step-arithmetic-word-problem-r20260728-155709.json in the repo.

Built by Sam Carlton