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%
23.0 KB · 577 lines yaml
Raw Blame History
1# Benchmark definitions (entities of type `benchmark`). Results come from connectors; definitions are curated with sources.2#3# Canonical fields (see aiatlas.ontology.benchmarks):4#   family            benchmark family key (swe-bench, gpqa, livebench…); `variant` names the member, `version` the dataset revision when one exists5#   family_head       true on the representative entry of a family — every other member gets a `variant_of` relation to it (aia seed)6#   metric            canonical metric key from ontology.benchmarks.METRICS (accuracy, pass@1, resolved, elo, index…); `metric_label` = human label7#   higher_is_better  score direction · metric_min / metric_max = bounds of the reported metric (null = unbounded)8#   harness           who runs the evaluation / with what scaffold; comparability_note = what makes two scores on this board (not) comparable9#   aliases           labels used by Artificial Analysis, LiveBench, model cards, papers… (the API resolves benchmarks by alias)10benchmarks:11  - key: mmlu12    name: MMLU13    aliases: [Massive Multitask Language Understanding, MMLU 5-shot, MMLU (5-shot)]14    family: mmlu15    variant: original16    family_head: true17    category: knowledge18    task: 57-subject multiple-choice questions19    metric: accuracy20    metric_label: accuracy21    unit: "%"22    metric_min: 023    metric_max: 10024    higher_is_better: true25    harness: self-reported by labs (shot count and prompt format vary)26    comparability_note: Widespread training-data contamination; 0-shot vs 5-shot and chain-of-thought prompting change scores by several points.27    creator: Hendrycks et al.28    website: https://github.com/hendrycks/test29    paper: https://arxiv.org/abs/2009.0330030    known_limitations: Widespread training-data contamination; saturated at the frontier.31    source_url: https://arxiv.org/abs/2009.0330032  - key: mmlu-pro33    name: MMLU-Pro34    aliases: [MMLU Pro, MMLU-Pro (0-shot CoT)]35    family: mmlu36    variant: Pro37    category: knowledge38    task: harder, 10-option MMLU variant39    metric: accuracy40    metric_label: accuracy41    unit: "%"42    metric_min: 043    metric_max: 10044    higher_is_better: true45    harness: self-reported by labs or independent evaluators46    comparability_note: 10-option questions; chain-of-thought is the standard regime, direct-answer scores are lower and not comparable.47    website: https://github.com/TIGER-AI-Lab/MMLU-Pro48    paper: https://arxiv.org/abs/2406.0157449    source_url: https://arxiv.org/abs/2406.0157450  - key: gpqa51    name: GPQA52    aliases: [GPQA main, Graduate-Level Google-Proof Q&A]53    family: gpqa54    variant: main55    family_head: true56    category: reasoning57    task: graduate-level science questions (main set, 448 questions)58    metric: accuracy59    metric_label: accuracy60    unit: "%"61    metric_min: 062    metric_max: 10063    higher_is_better: true64    harness: self-reported by labs or independent evaluators65    comparability_note: Main and Diamond subsets are different question sets; reasoning effort and sampling (pass@1 vs majority vote) change scores.66    website: https://github.com/idavidrein/gpqa67    paper: https://arxiv.org/abs/2311.1202268    source_url: https://arxiv.org/abs/2311.1202269  - key: gpqa-diamond70    name: GPQA Diamond71    aliases: [GPQA-Diamond, gpqa_diamond, GPQA (Diamond), GPQA Diamond (0-shot)]72    family: gpqa73    variant: Diamond74    category: reasoning75    task: graduate-level science questions — the 198-question Diamond subset (expert-validated, non-expert-failed)76    metric: accuracy77    metric_label: accuracy78    unit: "%"79    metric_min: 080    metric_max: 10081    higher_is_better: true82    harness: Artificial Analysis runs its own evaluation; labs self-report with varying sampling83    comparability_note: 198 questions → ±3 pt noise; reasoning effort, thinking budget and repeated sampling strongly affect scores.84    website: https://github.com/idavidrein/gpqa85    paper: https://arxiv.org/abs/2311.1202286    source_url: https://arxiv.org/abs/2311.1202287  - key: humaneval88    name: HumanEval89    aliases: [HumanEval pass@1, OpenAI HumanEval]90    family: humaneval91    variant: original92    family_head: true93    category: coding94    task: Python function synthesis from docstrings95    metric: pass@196    metric_label: pass@197    unit: "%"98    metric_min: 099    metric_max: 100100    higher_is_better: true101    harness: self-reported by labs102    comparability_note: Saturated; pass@1 vs pass@k and prompt wrapping differ across reporters.103    website: https://github.com/openai/human-eval104    paper: https://arxiv.org/abs/2107.03374105    known_limitations: Saturated; small (164 problems).106    source_url: https://arxiv.org/abs/2107.03374107  - key: swe-bench-verified108    name: SWE-bench Verified109    aliases: [SWE-bench, SWE-Bench Verified, SWE-bench-Verified, swebench verified, SWE Bench Verified]110    family: swe-bench111    variant: Verified112    category: coding113    task: resolve real GitHub issues (500 human-validated instances)114    metric: resolved115    metric_label: "% resolved"116    unit: "%"117    metric_min: 0118    metric_max: 100119    higher_is_better: true120    harness: submitter's agent scaffold (config.system); evaluation by the SWE-bench harness, submissions optionally checked by the SWE-bench team121    comparability_note: Scores depend on the agent scaffold as much as on the model; compare only within one scaffold (config.system) and one attempt regime.122    website: https://www.swebench.com123    paper: https://arxiv.org/abs/2310.06770124    known_limitations: Scaffold/agent dependent; results are not comparable across harnesses.125    source_url: https://www.swebench.com/126  - key: aider-polyglot127    name: Aider polyglot128    aliases: [Aider polyglot coding leaderboard, Aider Polyglot, aider polyglot benchmark, Aider]129    family: aider-polyglot130    variant: polyglot131    family_head: true132    category: coding133    task: 225 Exercism exercises in 6 languages, edit-format aware134    metric: pass_rate_2135    metric_label: pass rate (2 attempts)136    unit: "%"137    metric_min: 0138    metric_max: 100139    higher_is_better: true140    harness: aider (config.edit_format = whole | diff | diff-fenced | architect), second attempt after test feedback141    comparability_note: Edit format and aider version change results; the well-formed rate is a separate benchmark entry.142    website: https://aider.chat/docs/leaderboards/143    known_limitations: Depends on aider's edit format and prompting; cost column depends on provider pricing.144    source_url: https://aider.chat/docs/leaderboards/145  - key: aider-polyglot-well-formed146    name: Aider polyglot — well-formed responses147    aliases: [Aider well-formed rate, percent_cases_well_formed, Aider polyglot well-formed]148    family: aider-polyglot149    variant: well-formed150    category: coding151    task: share of aider polyglot cases where every edit was syntactically well formed (parsable edit blocks)152    metric: percent_cases_well_formed153    metric_label: "% cases well formed"154    unit: "%"155    metric_min: 0156    metric_max: 100157    higher_is_better: true158    harness: aider (config.edit_format), same runs as aider-polyglot159    comparability_note: Measures edit-format compliance, not correctness; only comparable within one edit format.160    website: https://aider.chat/docs/leaderboards/161    source_url: https://aider.chat/docs/leaderboards/162  - key: livebench163    name: LiveBench164    aliases: [LiveBench global average, LiveBench Global, LiveBench overall]165    family: livebench166    variant: global167    family_head: true168    category: general169    task: contamination-limited, monthly refreshed questions across 7 categories — overall mean of the category averages170    metric: global_average171    metric_label: global average172    unit: "%"173    metric_min: 0174    metric_max: 100175    higher_is_better: true176    harness: LiveBench official runs; config.release = question-set release date177    comparability_note: Each release is a new question set — compare scores within one release (run group) only; effort variants are configurations.178    website: https://livebench.ai179    paper: https://arxiv.org/abs/2406.19314180    source_url: https://livebench.ai/181  - key: livebench-reasoning182    name: LiveBench Reasoning183    aliases: [LiveBench — Reasoning, LiveBench reasoning average]184    family: livebench185    variant: Reasoning186    category: reasoning187    task: LiveBench reasoning category — mean of its subtasks (web of lies, zebra puzzles, spatial…)188    metric: average score189    metric_label: category average190    unit: "%"191    metric_min: 0192    metric_max: 100193    higher_is_better: true194    harness: LiveBench official runs195    comparability_note: Compare within one release only; subtask mix changes between releases.196    website: https://livebench.ai197    source_url: https://livebench.ai/198  - key: livebench-coding199    name: LiveBench Coding200    aliases: [LiveBench — Coding, LiveBench coding average]201    family: livebench202    variant: Coding203    category: coding204    task: LiveBench coding category — mean of its subtasks (LCB generation, code completion)205    metric: average score206    metric_label: category average207    unit: "%"208    metric_min: 0209    metric_max: 100210    higher_is_better: true211    harness: LiveBench official runs212    comparability_note: Compare within one release only; subtask mix changes between releases.213    website: https://livebench.ai214    source_url: https://livebench.ai/215  - key: livebench-agentic-coding216    name: LiveBench Agentic Coding217    aliases: [LiveBench — Agentic Coding, LiveBench agentic coding average]218    family: livebench219    variant: Agentic Coding220    category: coding221    task: LiveBench agentic coding category — mean of its subtasks222    metric: average score223    metric_label: category average224    unit: "%"225    metric_min: 0226    metric_max: 100227    higher_is_better: true228    harness: LiveBench official runs229    comparability_note: Compare within one release only; subtask mix changes between releases.230    website: https://livebench.ai231    source_url: https://livebench.ai/232  - key: livebench-mathematics233    name: LiveBench Mathematics234    aliases: [LiveBench — Mathematics, LiveBench math average, LiveBench Math]235    family: livebench236    variant: Mathematics237    category: math238    task: LiveBench mathematics category — mean of its subtasks (competition math, AMPS hard, olympiad)239    metric: average score240    metric_label: category average241    unit: "%"242    metric_min: 0243    metric_max: 100244    higher_is_better: true245    harness: LiveBench official runs246    comparability_note: Compare within one release only; subtask mix changes between releases.247    website: https://livebench.ai248    source_url: https://livebench.ai/249  - key: livebench-data-analysis250    name: LiveBench Data Analysis251    aliases: [LiveBench — Data Analysis, LiveBench data analysis average]252    family: livebench253    variant: Data Analysis254    category: general255    task: LiveBench data analysis category — mean of its subtasks (table reformatting, join, column type)256    metric: average score257    metric_label: category average258    unit: "%"259    metric_min: 0260    metric_max: 100261    higher_is_better: true262    harness: LiveBench official runs263    comparability_note: Compare within one release only; subtask mix changes between releases.264    website: https://livebench.ai265    source_url: https://livebench.ai/266  - key: livebench-language267    name: LiveBench Language268    aliases: [LiveBench — Language, LiveBench language average]269    family: livebench270    variant: Language271    category: general272    task: LiveBench language category — mean of its subtasks (typos, connections, plot unscrambling)273    metric: average score274    metric_label: category average275    unit: "%"276    metric_min: 0277    metric_max: 100278    higher_is_better: true279    harness: LiveBench official runs280    comparability_note: Compare within one release only; subtask mix changes between releases.281    website: https://livebench.ai282    source_url: https://livebench.ai/283  - key: livebench-if284    name: LiveBench Instruction Following285    aliases: [LiveBench — IF, LiveBench IF, LiveBench instruction following average]286    family: livebench287    variant: IF288    category: instruction-following289    task: LiveBench instruction-following category — mean of its subtasks (paraphrase, simplify, story generation, summarize)290    metric: average score291    metric_label: category average292    unit: "%"293    metric_min: 0294    metric_max: 100295    higher_is_better: true296    harness: LiveBench official runs297    comparability_note: Compare within one release only; subtask mix changes between releases.298    website: https://livebench.ai299    source_url: https://livebench.ai/300  - key: math-500301    name: MATH-500302    aliases: [MATH, MATH 500, MATH500]303    family: math304    variant: "500"305    family_head: true306    category: math307    task: competition mathematics (500-problem subset of MATH used since the PRM800K paper)308    metric: accuracy309    metric_label: accuracy310    unit: "%"311    metric_min: 0312    metric_max: 100313    higher_is_better: true314    harness: self-reported by labs or independent evaluators315    comparability_note: Saturated at the frontier; sampling regime (pass@1 vs majority vote) must match.316    paper: https://arxiv.org/abs/2103.03874317    source_url: https://arxiv.org/abs/2103.03874318  - key: aime-2024319    name: AIME 2024320    aliases: [AIME24, AIME 24, AIME I/II 2024]321    family: aime322    variant: "2024"323    version: "2024"324    category: math325    task: American Invitational Mathematics Examination 2024 problems (30 problems, I + II)326    metric: accuracy327    metric_label: accuracy328    unit: "%"329    metric_min: 0330    metric_max: 100331    higher_is_better: true332    harness: self-reported by labs or independent evaluators333    comparability_note: 30 problems → high variance; pass@1 averaged over samples vs majority voting (cons@64) are not comparable; 2024 problems appear in training data of later models.334    source_url: https://maa.org/maa-invitational-competitions/335  - key: aime-2025336    name: AIME 2025337    aliases: [AIME, AIME25, AIME 25, AIME I/II 2025]338    family: aime339    variant: "2025"340    version: "2025"341    family_head: true342    category: math343    task: American Invitational Mathematics Examination 2025 problems (30 problems, I + II)344    metric: accuracy345    metric_label: accuracy346    unit: "%"347    metric_min: 0348    metric_max: 100349    higher_is_better: true350    harness: Artificial Analysis runs its own evaluation; labs self-report with varying sampling351    comparability_note: 30 problems → high variance; pass@1 averaged over samples vs majority voting are not comparable; reasoning effort matters.352    known_limitations: 30 problems per year; high variance; often reported with majority voting.353    source_url: https://maa.org/maa-invitational-competitions/354  - key: arc-agi355    name: ARC-AGI356    aliases: [ARC-AGI-1, ARC AGI 1, Abstraction and Reasoning Corpus, ARC-AGI v1]357    family: arc-agi358    variant: "1"359    version: "1"360    family_head: true361    category: reasoning362    task: novel visual abstraction puzzles (ARC-AGI-1 semi-private / public evaluation sets)363    metric: accuracy364    metric_label: accuracy365    unit: "%"366    metric_min: 0367    metric_max: 100368    higher_is_better: true369    harness: ARC Prize Foundation verified runs (semi-private set) or self-reported public-set runs370    comparability_note: Public, semi-private and private evaluation sets differ; the ARC Prize reports cost per task alongside the score.371    website: https://arcprize.org372    source_url: https://arcprize.org/373  - key: arc-agi-2374    name: ARC-AGI-2375    aliases: [ARC AGI 2, ARC-AGI v2, ARC-AGI 2]376    family: arc-agi377    variant: "2"378    version: "2"379    category: reasoning380    task: harder abstraction puzzles released in 2025 (ARC-AGI-2), designed to resist brute-force program search381    metric: accuracy382    metric_label: accuracy383    unit: "%"384    metric_min: 0385    metric_max: 100386    higher_is_better: true387    harness: ARC Prize Foundation verified runs (semi-private set) or self-reported public-set runs388    comparability_note: Not comparable with ARC-AGI-1 scores; evaluation set (public / semi-private) and cost budget must match.389    website: https://arcprize.org/arc-agi/2/390    source_url: https://arcprize.org/arc-agi/2/391  - key: humanitys-last-exam392    name: Humanity's Last Exam393    aliases: [HLE, Humanity’s Last Exam, HLE (no tools)]394    family: humanitys-last-exam395    variant: full396    family_head: true397    category: knowledge398    task: expert-written frontier questions399    metric: accuracy400    metric_label: accuracy401    unit: "%"402    metric_min: 0403    metric_max: 100404    higher_is_better: true405    harness: Artificial Analysis and labs; text-only subset vs full multimodal set, with or without tools406    comparability_note: Tool use (search) and the text-only subset change scores substantially; check config before comparing.407    website: https://lastexam.ai408    paper: https://arxiv.org/abs/2501.14249409    source_url: https://lastexam.ai/410  - key: lmarena-text411    name: LMArena text leaderboard412    aliases: [Chatbot Arena, LMSYS Chatbot Arena, Arena Elo, LMArena, Arena Score]413    family: lmarena414    variant: text415    family_head: true416    category: preference417    task: crowdsourced pairwise human preference418    metric: elo419    metric_label: Elo / Bradley–Terry score420    unit: ""421    metric_min: 0422    metric_max: null423    higher_is_better: true424    harness: LMArena live human votes (Bradley–Terry model)425    comparability_note: Relative score that drifts as the model pool changes; only a snapshot of the same date is comparable; style control changes rankings.426    website: https://lmarena.ai427    known_limitations: Style bias; sampling of prompts by users.428    source_url: https://lmarena.ai/429  - key: mmmu430    name: MMMU431    aliases: [MMMU (val), Massive Multi-discipline Multimodal Understanding]432    family: mmmu433    variant: original434    family_head: true435    category: multimodal436    task: college-level multimodal understanding437    metric: accuracy438    metric_label: accuracy439    unit: "%"440    metric_min: 0441    metric_max: 100442    higher_is_better: true443    harness: self-reported by labs (validation split)444    comparability_note: Validation vs test split and chain-of-thought prompting differ across reporters.445    website: https://mmmu-benchmark.github.io446    paper: https://arxiv.org/abs/2311.16502447    source_url: https://arxiv.org/abs/2311.16502448  - key: tau-bench449    name: τ-bench450    aliases: [tau-bench, TAU-bench, tau bench, τ-Bench]451    family: tau-bench452    variant: v1453    family_head: true454    category: agentic455    task: tool-agent-user interaction in retail/airline domains456    metric: pass^1457    metric_label: pass^1458    unit: "%"459    metric_min: 0460    metric_max: 100461    higher_is_better: true462    harness: simulated user (LLM) + tool environment; pass^k over k trials463    comparability_note: Domain (retail / airline), pass^k regime and the user-simulator model must match.464    paper: https://arxiv.org/abs/2406.12045465    source_url: https://arxiv.org/abs/2406.12045466  - key: terminal-bench467    name: Terminal-Bench468    aliases: [TerminalBench, Terminal Bench, Terminal-Bench 1.0]469    family: terminal-bench470    variant: "1.0"471    version: "1.0"472    family_head: true473    category: agentic474    task: terminal tasks solved by agents475    metric: accuracy476    metric_label: accuracy477    unit: "%"478    metric_min: 0479    metric_max: 100480    higher_is_better: true481    harness: Terminus agent (official) or the submitter's agent; Artificial Analysis reports its own harness versions (config.variant)482    comparability_note: Agent scaffold and harness version change results; config.variant (v2.1 / v4.0 / hard) must match.483    website: https://www.tbench.ai484    source_url: https://www.tbench.ai/485  - key: terminal-bench-2486    name: Terminal-Bench 2.0487    aliases: [Terminal Bench 2, TerminalBench 2.0, Terminal-Bench 2, TB2]488    family: terminal-bench489    variant: "2.0"490    version: "2.0"491    category: agentic492    task: revised, harder set of terminal tasks (Terminal-Bench 2.0)493    metric: accuracy494    metric_label: accuracy495    unit: "%"496    metric_min: 0497    metric_max: 100498    higher_is_better: true499    harness: Terminus 2 agent (official) or the submitter's agent500    comparability_note: Not comparable with Terminal-Bench 1.0; agent scaffold must match.501    website: https://www.tbench.ai502    source_url: https://www.tbench.ai/503  - key: ifeval504    name: IFEval505    aliases: [IF-Eval, Instruction-Following Eval, IFEval strict]506    family: ifeval507    variant: original508    family_head: true509    category: instruction-following510    task: verifiable instruction following511    metric: accuracy512    metric_label: prompt-level strict accuracy513    unit: "%"514    metric_min: 0515    metric_max: 100516    higher_is_better: true517    harness: self-reported by labs518    comparability_note: Prompt-level vs instruction-level and strict vs loose accuracy are four different numbers.519    paper: https://arxiv.org/abs/2311.07911520    source_url: https://arxiv.org/abs/2311.07911521  - key: mteb522    name: MTEB523    aliases: [Massive Text Embedding Benchmark, MTEB (English), MTEB v1]524    family: mteb525    variant: v1526    family_head: true527    category: embeddings528    task: embedding tasks across retrieval, classification, clustering…529    metric: mean score530    metric_label: mean score531    unit: ""532    metric_min: 0533    metric_max: 100534    higher_is_better: true535    harness: MTEB library, official leaderboard536    comparability_note: Task set (English v1, v2, multilingual) and aggregation (mean of tasks vs mean of task types) must match.537    website: https://huggingface.co/spaces/mteb/leaderboard538    paper: https://arxiv.org/abs/2210.07316539    source_url: https://arxiv.org/abs/2210.07316540  - key: mmteb541    name: MMTEB542    aliases: [Massive Multilingual Text Embedding Benchmark, MTEB multilingual, MTEB (Multilingual)]543    family: mteb544    variant: MMTEB545    category: embeddings546    task: multilingual extension of MTEB (500+ tasks across 250+ languages)547    metric: mean score548    metric_label: mean score549    unit: ""550    metric_min: 0551    metric_max: 100552    higher_is_better: true553    harness: MTEB library, official leaderboard554    comparability_note: Not comparable with English-only MTEB; the leaderboard's Borda rank and task-type mean are different aggregations.555    website: https://huggingface.co/spaces/mteb/leaderboard556    paper: https://arxiv.org/abs/2502.13595557    source_url: https://arxiv.org/abs/2502.13595558  - key: artificial-analysis-intelligence-index559    name: Artificial Analysis Intelligence Index560    aliases: [AA Intelligence Index, Intelligence Index, Artificial Analysis Index]561    family: artificial-analysis-intelligence-index562    variant: index563    family_head: true564    category: composite565    task: composite of several evaluations run by Artificial Analysis566    metric: index567    metric_label: index568    unit: ""569    metric_min: 0570    metric_max: 100571    higher_is_better: true572    harness: Artificial Analysis internal runs; config.version = index methodology version573    comparability_note: Methodology versions change the component set — compare within one version (run group) only; reasoning-effort variants are configurations.574    website: https://artificialanalysis.ai575    known_limitations: Proprietary composite; methodology versions change.576    source_url: https://artificialanalysis.ai/methodology577