Terminal-Bench-Science
Lumina VerifiedLumina checked the published sources, versions, model labels, score extraction and conversion. External evaluations were run by their publishers; Lumina has not independently rerun or reproduced them.Accuracy
Terminal-Bench-Science
Lumina Score vs Cost per Task
Published per-test USD and Duration/Test retain their native definitions. Tokens use published averages or original aggregate counts divided by token_tasks_measured; total = input + output, with cache and reasoning subsets retained separately rather than counted twice. No task values are inferred from token prices or compute budgets.
Lumina Score vs Tokens
Published per-test USD and Duration/Test retain their native definitions. Tokens use published averages or original aggregate counts divided by token_tasks_measured; total = input + output, with cache and reasoning subsets retained separately rather than counted twice. No task values are inferred from token prices or compute budgets.
Lumina Score vs Time per Task
Published per-test USD and Duration/Test retain their native definitions. Tokens use published averages or original aggregate counts divided by token_tasks_measured; total = input + output, with cache and reasoning subsets retained separately rather than counted twice. No task values are inferred from token prices or compute budgets.
Published per-test USD and Duration/Test retain their native definitions. Tokens use published averages or original aggregate counts divided by token_tasks_measured; total = input + output, with cache and reasoning subsets retained separately rather than counted twice. No task values are inferred from token prices or compute budgets. Missing metrics are omitted from their graph; quality scores remain available.
Results table
Method & coverage
70 scientific terminal tasks with text bash or tmux-screen output. Mini-SWE-agent and Terminus 2 harnesses kept distinct; pass@1, infrastructure errors may be re-attempted, graded failures are not. Not directly comparable with native-agent three-trial official board.
Published per-test USD and Duration/Test retain their native definitions. Tokens use published averages or original aggregate counts divided by token_tasks_measured; total = input + output, with cache and reasoning subsets retained separately rather than counted twice. No task values are inferred from token prices or compute budgets.
How the Lumina conversion works
1,000 + 75 × ((original score − 14.285772727272722) / 16.559320477471893). The result is rounded once. Scores may tie.
Frozen initial observed text-only cohort; one mean per exact publisher release/API, balancing repeated settings/routes within that release. Original score and protocol remain separate. No inferred variants or future refitting.
- Calibration
- vals-terminal-bench-science-z75-v1 · experimental
- Reference mean
- 14.285772727272722
- Population SD
- 16.559320477471893
- Coverage
- Standalone benchmark.
Identity, exclusions & attribution
70 scientific terminal tasks with text bash or tmux-screen output. Mini-SWE-agent and Terminus 2 harnesses kept distinct; pass@1, infrastructure errors may be re-attempted, graded failures are not. Not directly comparable with native-agent three-trial official board.
Publisher standard errors retained; no confidence level or interval inferred. Aggregate reporting scope is recorded in the current source-specific publication review.
Standalone text benchmark; no new combined index. Source identity/effort/protocol qualifications remain in Data and CSV.
All published reasoning settings, including Claude Max, are included by default. Display labels use Low and xHigh; source labels remain in the retained records.
Official source · Publication terms / evidence
Publisher terms. Scores and resource figures are attributed to the original publisher.