SPB Git forge

spb/ai-atlas

Public
41commits 1branches 0releases
4.6 MBsize
maindefault branch
12 days agolast push
HTML 77.2% TypeScript 10.5% Python 9.6% JavaScript 2.5%
6.9 KB · 182 lines yaml
Raw Blame History
1# Benchmarks reported by the leaderboard connectors (SWE-bench boards, Artificial Analysis component evaluations). Fragment of registry/benchmarks.yaml2# (same canonical fields: family / variant / version / family_head / metric / metric_min / metric_max / higher_is_better / harness / comparability_note).3benchmarks:4  - key: swe-bench-full5    name: SWE-bench (full test split)6    aliases: [SWE-bench Test, SWE-bench Full, SWE-bench (full), swebench full]7    family: swe-bench8    variant: full9    family_head: true10    category: coding11    task: resolve real GitHub issues (2,294 instances)12    metric: resolved13    metric_label: "% resolved"14    unit: "%"15    metric_min: 016    metric_max: 10017    higher_is_better: true18    harness: submitter's agent scaffold (config.system); evaluation by the SWE-bench harness19    comparability_note: Scaffold-dependent; compare within one scaffold and attempt regime only.20    website: https://www.swebench.com21    paper: https://arxiv.org/abs/2310.0677022    source_url: https://www.swebench.com/23  - key: swe-bench-lite24    name: SWE-bench Lite25    aliases: [SWE-Bench Lite, swebench lite]26    family: swe-bench27    variant: Lite28    category: coding29    task: resolve real GitHub issues (300-instance subset)30    metric: resolved31    metric_label: "% resolved"32    unit: "%"33    metric_min: 034    metric_max: 10035    higher_is_better: true36    harness: submitter's agent scaffold (config.system); evaluation by the SWE-bench harness37    comparability_note: Scaffold-dependent; Lite is an easier subset — never compare with Verified or full scores.38    website: https://www.swebench.com39    paper: https://arxiv.org/abs/2310.0677040    known_limitations: Scaffold/agent dependent; results are not comparable across harnesses.41    source_url: https://www.swebench.com/42  - key: swe-bench-multimodal43    name: SWE-bench Multimodal44    aliases: [SWE-Bench Multimodal, SWE-bench MM]45    family: swe-bench46    variant: Multimodal47    category: coding48    task: resolve visual JavaScript issues (screenshots + code)49    metric: resolved50    metric_label: "% resolved"51    unit: "%"52    metric_min: 053    metric_max: 10054    higher_is_better: true55    harness: submitter's agent scaffold (config.system); evaluation by the SWE-bench harness56    comparability_note: Requires image input; scaffold-dependent.57    website: https://www.swebench.com/multimodal58    source_url: https://www.swebench.com/59  - key: swe-bench-multilingual60    name: SWE-bench Multilingual61    aliases: [SWE-Bench Multilingual]62    family: swe-bench63    variant: Multilingual64    category: coding65    task: resolve GitHub issues across 9 programming languages66    metric: resolved67    metric_label: "% resolved"68    unit: "%"69    metric_min: 070    metric_max: 10071    higher_is_better: true72    harness: submitter's agent scaffold (config.system); evaluation by the SWE-bench harness73    comparability_note: Scaffold-dependent; different instance set from Verified.74    website: https://www.swebench.com/multilingual75    source_url: https://www.swebench.com/76  - key: swe-bench-pro77    name: SWE-Bench Pro78    aliases: [SWE-bench Pro, SWE Bench Pro, swebench pro]79    family: swe-bench80    variant: Pro81    category: coding82    task: long-horizon, enterprise-grade software engineering tasks (public set) curated by Scale AI83    metric: resolved84    metric_label: "% resolved"85    unit: "%"86    metric_min: 087    metric_max: 10088    higher_is_better: true89    harness: Scale AI evaluation with a fixed scaffold per leaderboard column90    comparability_note: Public vs commercial (held-out) sets are different leaderboards; not comparable with SWE-bench Verified.91    website: https://scale.com/leaderboard/swe_bench_pro_public92    source_url: https://scale.com/leaderboard/swe_bench_pro_public93  - key: mmmu-pro94    name: MMMU-Pro95    aliases: [MMMU Pro]96    family: mmmu97    variant: Pro98    category: multimodal99    task: robust multimodal understanding (10-option, vision-only variants)100    metric: accuracy101    metric_label: accuracy102    unit: "%"103    metric_min: 0104    metric_max: 100105    higher_is_better: true106    harness: Artificial Analysis runs its own evaluation; labs self-report107    comparability_note: Standard (10-option) vs vision-only settings are different numbers.108    paper: https://arxiv.org/abs/2409.02813109    source_url: https://arxiv.org/abs/2409.02813110  - key: livecodebench111    name: LiveCodeBench112    aliases: [LCB, Live Code Bench]113    family: livecodebench114    variant: rolling115    family_head: true116    category: coding117    task: contamination-free competitive programming problems118    metric: pass@1119    metric_label: pass@1120    unit: "%"121    metric_min: 0122    metric_max: 100123    higher_is_better: true124    harness: Artificial Analysis runs its own evaluation; labs self-report on a chosen date window125    comparability_note: The problem window (release dates) differs between reporters — scores from different windows are not comparable.126    website: https://livecodebench.github.io127    paper: https://arxiv.org/abs/2403.07974128    source_url: https://arxiv.org/abs/2403.07974129  - key: scicode130    name: SciCode131    aliases: [SciCode benchmark]132    family: scicode133    variant: main134    family_head: true135    category: coding136    task: research-level scientific coding problems137    metric: accuracy138    metric_label: accuracy139    unit: "%"140    metric_min: 0141    metric_max: 100142    higher_is_better: true143    harness: Artificial Analysis runs its own evaluation144    comparability_note: Sub-problem vs main-problem accuracy are different numbers; with/without background prompts.145    website: https://scicode-bench.github.io146    paper: https://arxiv.org/abs/2407.13168147    source_url: https://arxiv.org/abs/2407.13168148  - key: ifbench149    name: IFBench150    aliases: [IF-Bench]151    family: ifeval152    variant: IFBench153    category: instruction-following154    task: precise instruction following with novel constraints155    metric: accuracy156    metric_label: accuracy157    unit: "%"158    metric_min: 0159    metric_max: 100160    higher_is_better: true161    harness: Artificial Analysis runs its own evaluation162    comparability_note: Different constraint set from IFEval; not comparable with IFEval scores.163    paper: https://arxiv.org/abs/2507.02833164    source_url: https://arxiv.org/abs/2507.02833165  - key: tau2-bench166    name: τ²-bench167    aliases: [tau2-bench, TAU2-bench, tau-squared bench, τ²-Bench, τ2-bench, τ²-Bench Telecom, tau2 bench telecom]168    family: tau-bench169    variant: τ²170    category: agentic171    task: dual-control tool-agent-user interaction (telecom, retail, airline)172    metric: pass^1173    metric_label: pass^1174    unit: "%"175    metric_min: 0176    metric_max: 100177    higher_is_better: true178    harness: simulated user (LLM) + dual-control tool environment; Artificial Analysis reports the Telecom domain179    comparability_note: Domain (config.variant = Telecom / Retail / Airline) and the user-simulator model must match.180    paper: https://arxiv.org/abs/2506.07982181    source_url: https://arxiv.org/abs/2506.07982182