import{r as F,j as a,y as le,L as ne,b as ot,h as ct,u as dt,d as ut,k as mt}from"./index-C2uVRBgm.js";import{R as ht}from"./ResearchFollowButton-bifNs9Eb.js";import{u as te}from"./useQuery-DceN3et-.js";import{c as re}from"./catalog-CUh96gFy.js";import{C as ft,E as pt,e as gt}from"./ExternalSourceEvalBadge-CguvmYcg.js";import{L as oe}from"./LatexTitle-D0kKUw3O.js";import{S as Ke}from"./switch-Cj-NJg9J.js";import{I as xt}from"./components-B2P9mD5v.js";import{S as ce}from"./skeleton-BtMQpCs1.js";import{A as bt,a as yt}from"./alert-7oLVYxe5.js";import{B as wt}from"./BenchmarkTrendSparkline-DB8DA6c6.js";import{s as vt}from"./safeExternalHref-CkBrsh84.js";import{s as jt}from"./slug-BpH-Nkvz.js";import"./useMutation-B2A2sn3L.js";import"./research-C0QOOVFw.js";import"./plus-S7qjFpAw.js";import"./papers-BSHjPS5U.js";import"./hover-card-C-7bnL__.js";import"./index-DowQVb_h.js";import"./paperQueryKeys-BP2E-0yf.js";import"./colors-CYJo8Vj2.js";function _t(e,t,r){const s=t?(e??[]).filter(l=>l.task_id===t):e??[],n=s.filter(l=>l.is_open===!1),i=n.reduce((l,c)=>l?(c.best_rank??Number.POSITIVE_INFINITY)<(l.best_rank??Number.POSITIVE_INFINITY)?c:l:c,null);return{visible:r?s.filter(l=>l.is_open!==!1):s,hiddenClosed:n,topHiddenClosed:i}}function ie(e){if(e==null)return null;if(typeof e=="number")return Number.isFinite(e)?e:null;const t=e.includes("±")?e.split("±")[0].trim():e.trim(),r=parseFloat(t);return Number.isFinite(r)?r:null}const Ge=[{label:"All",maxParametersExclusive:void 0},{label:"<10M",maxParametersExclusive:1e7},{label:"<100M",maxParametersExclusive:1e8},{label:"<1B",maxParametersExclusive:1e9},{label:"<3B",maxParametersExclusive:3e9},{label:"<6B",maxParametersExclusive:6e9},{label:"<12B",maxParametersExclusive:12e9},{label:"<32B",maxParametersExclusive:32e9},{label:"<128B",maxParametersExclusive:128e9},{label:"<500B",maxParametersExclusive:5e11}];function St(e){const t=e===null?void 0:Number(e);return Ge.some(r=>r.maxParametersExclusive===t)?t:void 0}function kt(e){return new Intl.NumberFormat("en",{notation:"compact",maximumSignificantDigits:3}).format(e)}function Ct(e){return[e.paper_id,e.task_id,e.dataset_id,e.model_name,e.harness??""].join("|")}function Et(e){const t=new Map;for(const r of e){const s=Ct(r),n=t.get(s);if(!n){t.set(s,{...r,metrics:{...r.metrics||{}},sourceEvaluationIds:[r.id],ranksByMetric:r.best_metric&&r.best_rank!=null?{[r.best_metric]:r.best_rank}:{},parameterCount:r.num_parameters??null,hasParameterConflict:!1});continue}n.metrics={...n.metrics,...r.metrics||{}},n.sourceEvaluationIds.push(r.id),r.best_metric&&r.best_rank!=null&&(n.ranksByMetric[r.best_metric]=r.best_rank),r.num_parameters!=null&&!n.hasParameterConflict&&(n.parameterCount===null?n.parameterCount=r.num_parameters:n.parameterCount!==r.num_parameters&&(n.parameterCount=null,n.hasParameterConflict=!0))}return Array.from(t.values())}function Le(e,t){return t===void 0?e:e.filter(r=>r.parameterCount!==null&&r.parameterCount<t)}function Nt(e,t){return t?e.id===t||e.sourceEvaluationIds.includes(t):!1}function Ie(e,t){return e.ranksByMetric[t]??e.best_rank??9999}function Rt(e){const t=Date.parse(e);return Number.isNaN(t)?"":new Date(t).toISOString().slice(0,10)}function Pt(e){const t=new Map;for(const s of e)s.evaluated_on&&t.set(s.evaluated_on,(t.get(s.evaluated_on)??0)+1);const r=Array.from(t.values()).reduce((s,n)=>s+n,0);for(const[s,n]of t)if(n>=2&&n*2>r)return s;return null}function $t(e,t){const r=e.evaluated_on;return r&&r!==t&&r!==Rt(e.updated_at)?r:e.paper_published_date||""}function Mt(e,t,r){if(e.length===0)return{points:[],sota:0,sotaModel:""};const s=Pt(e),n=e.map(u=>{var R;const h=ie((R=u.metrics)==null?void 0:R[t]);return{eval:u,value:h,date:$t(u,s)}}).filter(u=>u.value!==null&&!!u.date).sort((u,h)=>u.date.localeCompare(h.date)),i=[];let l=null;for(const u of n){const h=l==null?!0:r==="↓"?u.value<l:u.value>l;h&&(l=u.value),i.push({value:u.value,date:u.date,dateLabel:u.date.slice(0,7),modelName:u.eval.model_name||"",paperTitle:u.eval.paper_title||"",paperRouteId:u.eval.paper_arxiv_id||u.eval.paper_id||null,isSota:h})}const d=e.map(u=>{var h;return{eval:u,value:ie((h=u.metrics)==null?void 0:h[t])}}).filter(u=>u.value!==null).sort((u,h)=>r==="↓"?u.value-h.value:h.value-u.value)[0],f=(d==null?void 0:d.value)??0,x=(d==null?void 0:d.eval.model_name)||(d==null?void 0:d.eval.paper_title)||"";return{points:i,sota:f,sotaModel:x}}function Qe(e,t,r,s){const n=t.find(i=>i.id===s);if(e.length===0)return null;if(n!=null&&n.task_id&&e.some(i=>i.id===n.task_id))return n.task_id;if(r){const i=e.find(l=>l.slug===r);if(i)return i.id}if(e.length===1)return e[0].id;if(t.length>0){const i=new Map;for(const d of t)i.set(d.task_id,(i.get(d.task_id)||0)+1);let l=null,c=-1;for(const d of e){const f=i.get(d.id)||0;f>c&&(c=f,l=d.id)}if(l)return l}return e[0].id}function Tt(e){var _,X,Y;const{evaluations:t,tasks:r,taskSlug:s,highlightEvalId:n,parameterBound:i,showMoreMetrics:l,fittingMetricCount:c=1}=e,d=Qe(r,t,s,n),f=r.find(o=>o.id===d)||r[0]||null,x=t.find(o=>o.id===n),u=_t(t,d,e.hideClosed),h=Et(u.visible),R=Le(h,i),E=e.configuredMetrics.map(o=>o.metric_name),L=Array.from(new Set(h.flatMap(o=>[...o.best_metric?[o.best_metric]:[],...Object.keys(o.metrics||{})]))),S=E.length?E:L,g=new Map;for(const o of h){const m=o.best_metric;!m||ie(o.metrics[m])===null||E.length&&!S.includes(m)||g.set(m,(g.get(m)||0)+1)}const k=((_=[...g].sort((o,m)=>m[1]-o[1])[0])==null?void 0:_[0])||S[0]||"",j=S.find(o=>{var m;return o.toLocaleLowerCase()===((m=e.metric)==null?void 0:m.trim().toLocaleLowerCase())})||(x==null?void 0:x.best_metric)||k,D=((X=e.directions)==null?void 0:X.directions[j])==="lower_is_better"?"↓":"↑",W=j&&S.includes(j)?[j,...S.filter(o=>o!==j)]:S,w=W.length?l?W:W.slice(0,Math.max(1,c)):j?[j]:[],q=R.slice().sort((o,m)=>{const p=ie(o.metrics[j]),y=ie(m.metrics[j]);return p!==null&&y!==null?D==="↓"?p-y:y-p:p!==null?-1:y!==null?1:Ie(o,j)-Ie(m,j)});let I=null,N=0;const Q=q.map((o,m)=>{const p=ie(o.metrics[j]);return p!==null&&p!==I&&(I=p,N=m+1),{evaluation:o,rank:p===null?null:N,isHighlighted:Nt(o,n)}}),{points:z,sota:ee,sotaModel:O}=Mt(R,j,D),$=i===void 0?e.initialSummary:void 0,U=e.ready.evaluations&&e.ready.tasks&&e.ready.directions&&(!d||e.ready.metrics),K={parameter_coverage_known:h.filter(o=>o.parameterCount!==null).length,parameter_coverage_total:h.length};return{currentTaskId:d,primaryTask:f,activeMetric:j,arrow:D,allowedMetrics:S,orderedMetrics:W,visibleMetrics:w,rows:Q,chartPoints:z,highlightedEval:((Y=Q.find(o=>o.isHighlighted))==null?void 0:Y.evaluation)||null,hiddenClosedEvals:u.hiddenClosed,hiddenClosedTopEval:u.topHiddenClosed,parameterBound:i,parameterCoverage:K,hasParameterMetadata:K.parameter_coverage_known>0,parameterFilters:Ge.map(o=>({...o,selected:i===o.maxParametersExclusive,disabled:i!==o.maxParametersExclusive&&Le(h,o.maxParametersExclusive).length===0})),displayedMetric:U?j:$==null?void 0:$.metric,displayedSota:U?ee:$==null?void 0:$.value,displayedSotaModel:U?O:$==null?void 0:$.model_name,isSummaryLoading:!U&&!$,isInitialFilteredViewLoading:i!==void 0&&!U}}function Ft(e,t,r,s=1,n=0){const i=e-t.fixed-(r?t.harness:0);if(t.metric<=0)return 1;const l=Math.floor(i/t.metric);if(l<s)return Math.max(1,l);const c=Math.floor((i-n)/t.metric);return Math.max(1,s,c)}function he(e,t){return parseFloat(e.getPropertyValue(t))||0}function Lt(e){const t=getComputedStyle(e);return{fixed:["--bm-rank-width","--bm-model-width","--bm-parameters-width","--bm-paper-width","--bm-year-width"].reduce((r,s)=>r+he(t,s),0),harness:he(t,"--bm-harness-width"),metric:he(t,"--bm-metric-width")}}const It=20;function Ae(e,t){return e?a.jsx(oe,{title:e}):t||"—"}function At({model:e,showHarness:t,setMetric:r,highlightRowRef:s,onFittingMetricCountChange:n}){const{rows:i,visibleMetrics:l,activeMetric:c,arrow:d}=e,[f,x]=F.useState(null);return F.useEffect(()=>{if(!f)return;const u=()=>{f.classList.toggle("has-horizontal-overflow",f.scrollWidth>f.clientWidth);const R=f.firstElementChild;if(!R)return;const E=f.clientWidth,L=Lt(R),S=document.documentElement,g=S.scrollHeight>S.clientHeight?0:It;n(k=>Ft(E,L,t,k,g))},h=new ResizeObserver(u);return h.observe(f),f.firstElementChild&&h.observe(f.firstElementChild),u(),()=>h.disconnect()},[f,t,n]),a.jsx("div",{ref:x,className:"bm-table-scroll",role:"region","aria-label":"Leaderboard metrics",tabIndex:0,children:a.jsxs("table",{id:"benchmark-leaderboard",className:"bm-table",style:{"--bm-metric-count":l.length,"--bm-harness-count":t?1:0},children:[a.jsxs("colgroup",{children:[a.jsx("col",{className:"bm-width-rank"}),a.jsx("col",{className:"bm-width-model"}),a.jsx("col",{className:"bm-width-parameters"}),t&&a.jsx("col",{className:"bm-width-harness"}),l.map(u=>a.jsx("col",{className:"bm-width-metric"},u)),a.jsx("col",{}),a.jsx("col",{className:"bm-width-year"})]}),a.jsx("thead",{children:a.jsxs("tr",{children:[a.jsx("th",{className:"col-rank",children:"Rank"}),a.jsx("th",{className:"col-model",children:"Model"}),a.jsx("th",{className:"col-parameters",children:"Parameters"}),t&&a.jsx("th",{className:"col-harness",children:"Harness"}),l.map(u=>{const h=u===c;return a.jsx("th",{className:`col-value ${h?"is-active":""}`,"aria-sort":h?d==="↓"?"ascending":"descending":"none",children:a.jsxs("button",{type:"button",className:"bm-metric-header",onClick:()=>r(u),"aria-pressed":h,title:`Sort by ${u}`,children:[u,h?` ${d}`:""]})},u)}),a.jsx("th",{className:"col-paper",children:"Paper"}),a.jsx("th",{className:"col-year",children:"Year"})]})}),a.jsx("tbody",{children:i.map(({evaluation:u,rank:h,isHighlighted:R})=>{var S;const E=u.paper_arxiv_id||u.paper_id,L=u.paper_title||(u.paper_arxiv_id?`arXiv:${u.paper_arxiv_id}`:null);return a.jsxs("tr",{ref:R?s:void 0,className:R?"is-highlighted":void 0,children:[a.jsx("td",{className:"col-rank",children:h??"—"}),a.jsxs("td",{className:"col-model",children:[R&&a.jsx("span",{className:"bm-row-marker","aria-hidden":"true"}),u.hf_model_url?a.jsx("a",{href:u.hf_model_url,target:"_blank",rel:"noopener noreferrer",className:"bm-model-link",title:`Open ${u.model_name||"model"} on Hugging Face`,children:u.model_name||"—"}):u.model_name||"—",u.is_open===!1&&a.jsx(ft,{}),(u.external??u.is_external)&&a.jsx(pt,{sourceUrl:u.source_url||u.external_source_url||u.result_url})]}),a.jsx("td",{className:"col-parameters",title:u.parameterCount===null?void 0:u.parameterCount.toLocaleString("en-US"),"aria-label":u.parameterCount===null?"Parameters unknown":`${u.parameterCount.toLocaleString("en-US")} parameters`,children:u.parameterCount===null?"—":kt(u.parameterCount)}),t&&a.jsx("td",{className:"col-harness",children:u.harness||"—"}),l.map(g=>{var j;const k=g===c,B=(j=u.metrics)==null?void 0:j[g];return a.jsx("td",{className:`col-value${k?" is-active":""}${k&&h===1?" is-best":""}`,children:B!=null&&B!==""?le(B):"—"},g)}),a.jsx("td",{className:"col-paper",children:E?a.jsx(ne,{to:`/paper/${E}`,title:u.paper_title||L||void 0,children:Ae(u.paper_title,L)}):Ae(u.paper_title,L)}),a.jsx("td",{className:"col-year",children:((S=u.evaluated_on||u.paper_published_date)==null?void 0:S.slice(0,4))||"—"})]},u.id)})})]})})}function Be(e){const t=e.length===7?`${e}-01`:e,r=new Date(t).getTime();return Number.isFinite(r)?r:0}const fe=["Jan","Feb","Mar","Apr","May","Jun","Jul","Aug","Sep","Oct","Nov","Dec"];function pe(e){const t=new Date(e);return{year:t.getFullYear(),month:t.getMonth()}}function Bt(e,t){if(e===t){const{year:d,month:f}=pe(e);return[{timestamp:new Date(d,f,1).getTime(),key:`${d}-${f}`,label:`${fe[f]} ${d}`}]}const r=pe(e),s=pe(t);if(r.year===s.year){const d=s.month-r.month+1;let f=1;const x=12;d>x&&(f=Math.ceil(d/x));const u=[];for(let h=r.month;h<=s.month;h+=f)u.push({timestamp:new Date(r.year,h,1).getTime(),key:`${r.year}-${h}`,label:d<=3?`${fe[h]} ${r.year}`:fe[h]});return u}const n=s.year-r.year+1;let i=1;const l=12;n>l&&(i=Math.ceil(n/l));const c=[];for(let d=r.year;d<=s.year;d+=i)c.push({timestamp:new Date(d,0,1).getTime(),key:String(d),label:String(d)});return c}const Dt=({points:e,metric:t})=>{const r=ot(),[s,n]=F.useState(null),i=880,l=320,c=60,d=30,f=40,x=50,u=e.map(o=>o.value),h=Math.min(...u),E=Math.max(...u)-h||1,L=i-c-d,S=l-f-x,g=e.map(o=>Be(o.date)),k=Math.min(...g),B=Math.max(...g),j=B-k||1,D=o=>k===B?c+L/2:c+(o-k)/j*L,W=e.map((o,m)=>{const p=D(Be(o.date)),y=(o.value-h)/E,P=f+S*(1-y);return{x:p,y:P,data:o,idx:m}}),w=W.filter(o=>o.data.isSota);let q="";if(w.length>=2){q=`M${w[0].x},${w[0].y}`;for(let o=1;o<w.length;o++){const m=w[o-1],p=w[o],y=(m.x+p.x)/2;q+=` C${y},${m.y} ${y},${p.y} ${p.x},${p.y}`}}const I=5.4,N=8,Q=9,z=o=>{const m=o.data.modelName||o.data.paperTitle||"";if(!m)return null;const p=m.length>22?m.slice(0,21)+"…":m,y=p.length*I;let P=o.x-y/2,C=o.x+y/2,v="middle";return P<c?(P=c,C=c+y,v="start"):C>i-d&&(C=i-d,P=C-y,v="end"),{x:v==="start"?c:v==="end"?i-d:o.x,y:o.y-Q,text:p,anchor:v,left:P,right:C}},ee=[],O=w.length>0?z(w[w.length-1]):null;let $=-1/0;for(let o=0;o<w.length-1;o++){const m=z(w[o]);m&&(m.left<$+N||O&&m.right>O.left-N||($=m.right,ee.push({x:m.x,y:m.y,text:m.text,anchor:m.anchor})))}O&&ee.push({x:O.x,y:O.y,text:O.text,anchor:O.anchor});const U=Bt(k,B),K=5,_=o=>n(m=>m===o?null:m),X=o=>{o&&r(`/paper/${o}`)},Y=(o,m)=>{const y=`${o.data.modelName||o.data.paperTitle||"Unknown model"}: ${le(o.data.value)} on ${o.data.dateLabel}`,P=!!o.data.paperRouteId;return a.jsxs("g",{children:[m,a.jsx("circle",{cx:o.x,cy:o.y,r:"10",fill:"transparent",style:{cursor:P?"pointer":"default",outline:"none"},tabIndex:0,role:P?"link":"img","aria-label":y,onMouseEnter:()=>n(o.idx),onMouseLeave:()=>_(o.idx),onFocus:()=>n(o.idx),onBlur:()=>_(o.idx),onClick:()=>X(o.data.paperRouteId),onKeyDown:C=>{P&&(C.key==="Enter"||C.key===" ")&&(C.preventDefault(),X(o.data.paperRouteId))},children:a.jsx("title",{children:y})})]},o.idx)};return a.jsxs("svg",{width:"100%",viewBox:`0 0 ${i} ${l}`,style:{display:"block"},children:[a.jsx("text",{x:c-38,y:f+S/2,fontSize:"11",fontFamily:"var(--mono)",fill:"var(--ink-faint)",textAnchor:"middle",transform:`rotate(-90 ${c-38} ${f+S/2})`,children:t}),Array.from({length:K}).map((o,m)=>{const p=m/(K-1),y=f+S*(1-p),P=h+E*p;return a.jsxs("g",{children:[a.jsx("line",{x1:c,y1:y,x2:i-d,y2:y,stroke:"var(--hairline)",strokeWidth:"1"}),a.jsx("text",{x:c-10,y:y+4,fontSize:"11",fontFamily:"var(--mono)",fill:"var(--ink-faint)",textAnchor:"end",children:le(P)})]},m)}),U.map(o=>{const m=D(o.timestamp);return a.jsx("text",{x:m,y:l-x+18,fontSize:"11",fontFamily:"var(--mono)",fill:"var(--ink-soft)",textAnchor:"middle",children:o.label},o.key)}),W.filter(o=>!o.data.isSota).map(o=>{const m=s===o.idx;return Y(o,a.jsx("circle",{cx:o.x,cy:o.y,r:m?3:2,fill:"var(--ink-faint)",opacity:m?1:.45,style:{transition:"r 0.12s ease, opacity 0.12s ease"}}))}),q&&a.jsx("path",{d:q,stroke:"var(--accent)",strokeWidth:"2",fill:"none",strokeLinejoin:"round",strokeLinecap:"round"}),w.map(o=>{const m=s===o.idx;return Y(o,a.jsx("circle",{cx:o.x,cy:o.y,r:m?4.5:3,fill:m?"var(--accent)":"var(--bg-elevated)",stroke:"var(--accent)",strokeWidth:"1.5",style:{transition:"r 0.12s ease, fill 0.12s ease"}}))}),ee.map((o,m)=>a.jsx("text",{x:o.x,y:o.y,fontSize:"9.5",fontFamily:"var(--mono)",fill:"var(--ink-soft)",textAnchor:o.anchor,pointerEvents:"none",style:{paintOrder:"stroke",stroke:"var(--bg-elevated)",strokeWidth:3,strokeLinejoin:"round"},children:o.text},`label-${m}`)),s!==null&&W[s]&&(()=>{const o=W[s],m=o.data.modelName||o.data.paperTitle||"Unknown model",y=`${m.length>40?m.slice(0,39)+"…":m}: ${le(o.data.value)}`,C=Math.max(120,Math.min(340,y.length*6.6+24)),v=30;let H=o.x-C/2;H<c&&(H=c),H+C>i-d&&(H=i-d-C);const G=o.y-v-14>=f?o.y-v-12:o.y+12;return a.jsxs("g",{pointerEvents:"none",children:[a.jsx("rect",{x:H,y:G,width:C,height:v,rx:"6",ry:"6",fill:"var(--bg-elevated)",stroke:"var(--hairline-strong)",strokeWidth:"1",style:{filter:"drop-shadow(0 2px 6px rgba(0,0,0,0.08))"}}),a.jsx("text",{x:H+12,y:G+19,fontSize:"12",fontFamily:"var(--sans)",fill:"var(--ink)",fontWeight:500,children:y})]})})()]})};function Ht(e,t){if(e.match(/^[a-z]+:\/\//i))return e;if(e.match(/^\/\//))return window.location.protocol+e;if(e.match(/^[a-z]+:/i))return e;const r=document.implementation.createHTMLDocument(),s=r.createElement("base"),n=r.createElement("a");return r.head.appendChild(s),r.body.appendChild(n),t&&(s.href=t),n.href=e,n.href}const Vt=(()=>{let e=0;const t=()=>`0000${(Math.random()*36**4<<0).toString(36)}`.slice(-4);return()=>(e+=1,`u${t()}${e}`)})();function Z(e){const t=[];for(let r=0,s=e.length;r<s;r++)t.push(e[r]);return t}let ae=null;function Xe(e={}){return ae||(e.includeStyleProperties?(ae=e.includeStyleProperties,ae):(ae=Z(window.getComputedStyle(document.documentElement)),ae))}function de(e,t){const s=(e.ownerDocument.defaultView||window).getComputedStyle(e).getPropertyValue(t);return s?parseFloat(s.replace("px","")):0}function Wt(e){const t=de(e,"border-left-width"),r=de(e,"border-right-width");return e.clientWidth+t+r}function Ot(e){const t=de(e,"border-top-width"),r=de(e,"border-bottom-width");return e.clientHeight+t+r}function Ye(e,t={}){const r=t.width||Wt(e),s=t.height||Ot(e);return{width:r,height:s}}function Ut(){let e,t;try{t=process}catch{}const r=t&&t.env?t.env.devicePixelRatio:null;return r&&(e=parseInt(r,10),Number.isNaN(e)&&(e=1)),e||window.devicePixelRatio||1}const V=16384;function qt(e){(e.width>V||e.height>V)&&(e.width>V&&e.height>V?e.width>e.height?(e.height*=V/e.width,e.width=V):(e.width*=V/e.height,e.height=V):e.width>V?(e.height*=V/e.width,e.width=V):(e.width*=V/e.height,e.height=V))}function zt(e,t={}){return e.toBlob?new Promise(r=>{e.toBlob(r,t.type?t.type:"image/png",t.quality?t.quality:1)}):new Promise(r=>{const s=window.atob(e.toDataURL(t.type?t.type:void 0,t.quality?t.quality:void 0).split(",")[1]),n=s.length,i=new Uint8Array(n);for(let l=0;l<n;l+=1)i[l]=s.charCodeAt(l);r(new Blob([i],{type:t.type?t.type:"image/png"}))})}function ue(e){return new Promise((t,r)=>{const s=new Image;s.onload=()=>{s.decode().then(()=>{requestAnimationFrame(()=>t(s))})},s.onerror=r,s.crossOrigin="anonymous",s.decoding="async",s.src=e})}async function Kt(e){return Promise.resolve().then(()=>new XMLSerializer().serializeToString(e)).then(encodeURIComponent).then(t=>`data:image/svg+xml;charset=utf-8,${t}`)}async function Gt(e,t,r){const s="http://www.w3.org/2000/svg",n=document.createElementNS(s,"svg"),i=document.createElementNS(s,"foreignObject");return n.setAttribute("width",`${t}`),n.setAttribute("height",`${r}`),n.setAttribute("viewBox",`0 0 ${t} ${r}`),i.setAttribute("width","100%"),i.setAttribute("height","100%"),i.setAttribute("x","0"),i.setAttribute("y","0"),i.setAttribute("externalResourcesRequired","true"),n.appendChild(i),i.appendChild(e),Kt(n)}const A=(e,t)=>{if(e instanceof t)return!0;const r=Object.getPrototypeOf(e);return r===null?!1:r.constructor.name===t.name||A(r,t)};function Qt(e){const t=e.getPropertyValue("content");return`${e.cssText} content: '${t.replace(/'|"/g,"")}';`}function Xt(e,t){return Xe(t).map(r=>{const s=e.getPropertyValue(r),n=e.getPropertyPriority(r);return`${r}: ${s}${n?" !important":""};`}).join(" ")}function Yt(e,t,r,s){const n=`.${e}:${t}`,i=r.cssText?Qt(r):Xt(r,s);return document.createTextNode(`${n}{${i}}`)}function De(e,t,r,s){const n=window.getComputedStyle(e,r),i=n.getPropertyValue("content");if(i===""||i==="none")return;const l=Vt();try{t.className=`${t.className} ${l}`}catch{return}const c=document.createElement("style");c.appendChild(Yt(l,r,n,s)),t.appendChild(c)}function Jt(e,t,r){De(e,t,":before",r),De(e,t,":after",r)}const He="application/font-woff",Ve="image/jpeg",Zt={woff:He,woff2:He,ttf:"application/font-truetype",eot:"application/vnd.ms-fontobject",png:"image/png",jpg:Ve,jpeg:Ve,gif:"image/gif",tiff:"image/tiff",svg:"image/svg+xml",webp:"image/webp"};function er(e){const t=/\.([^./]*?)$/g.exec(e);return t?t[1]:""}function ye(e){const t=er(e).toLowerCase();return Zt[t]||""}function tr(e){return e.split(/,/)[1]}function xe(e){return e.search(/^(data:)/)!==-1}function rr(e,t){return`data:${t};base64,${e}`}async function Je(e,t,r){const s=await fetch(e,t);if(s.status===404)throw new Error(`Resource "${s.url}" not found`);const n=await s.blob();return new Promise((i,l)=>{const c=new FileReader;c.onerror=l,c.onloadend=()=>{try{i(r({res:s,result:c.result}))}catch(d){l(d)}},c.readAsDataURL(n)})}const ge={};function ar(e,t,r){let s=e.replace(/\?.*/,"");return r&&(s=e),/ttf|otf|eot|woff2?/i.test(s)&&(s=s.replace(/.*\//,"")),t?`[${t}]${s}`:s}async function we(e,t,r){const s=ar(e,t,r.includeQueryParams);if(ge[s]!=null)return ge[s];r.cacheBust&&(e+=(/\?/.test(e)?"&":"?")+new Date().getTime());let n;try{const i=await Je(e,r.fetchRequestInit,({res:l,result:c})=>(t||(t=l.headers.get("Content-Type")||""),tr(c)));n=rr(i,t)}catch(i){n=r.imagePlaceholder||"";let l=`Failed to fetch resource: ${e}`;i&&(l=typeof i=="string"?i:i.message),l&&console.warn(l)}return ge[s]=n,n}async function sr(e){const t=e.toDataURL();return t==="data:,"?e.cloneNode(!1):ue(t)}async function nr(e,t){if(e.currentSrc){const i=document.createElement("canvas"),l=i.getContext("2d");i.width=e.clientWidth,i.height=e.clientHeight,l==null||l.drawImage(e,0,0,i.width,i.height);const c=i.toDataURL();return ue(c)}const r=e.poster,s=ye(r),n=await we(r,s,t);return ue(n)}async function ir(e,t){var r;try{if(!((r=e==null?void 0:e.contentDocument)===null||r===void 0)&&r.body)return await me(e.contentDocument.body,t,!0)}catch{}return e.cloneNode(!1)}async function lr(e,t){return A(e,HTMLCanvasElement)?sr(e):A(e,HTMLVideoElement)?nr(e,t):A(e,HTMLIFrameElement)?ir(e,t):e.cloneNode(Ze(e))}const or=e=>e.tagName!=null&&e.tagName.toUpperCase()==="SLOT",Ze=e=>e.tagName!=null&&e.tagName.toUpperCase()==="SVG";async function cr(e,t,r){var s,n;if(Ze(t))return t;let i=[];return or(e)&&e.assignedNodes?i=Z(e.assignedNodes()):A(e,HTMLIFrameElement)&&(!((s=e.contentDocument)===null||s===void 0)&&s.body)?i=Z(e.contentDocument.body.childNodes):i=Z(((n=e.shadowRoot)!==null&&n!==void 0?n:e).childNodes),i.length===0||A(e,HTMLVideoElement)||await i.reduce((l,c)=>l.then(()=>me(c,r)).then(d=>{d&&t.appendChild(d)}),Promise.resolve()),t}function dr(e,t,r){const s=t.style;if(!s)return;const n=window.getComputedStyle(e);n.cssText?(s.cssText=n.cssText,s.transformOrigin=n.transformOrigin):Xe(r).forEach(i=>{let l=n.getPropertyValue(i);i==="font-size"&&l.endsWith("px")&&(l=`${Math.floor(parseFloat(l.substring(0,l.length-2)))-.1}px`),A(e,HTMLIFrameElement)&&i==="display"&&l==="inline"&&(l="block"),i==="d"&&t.getAttribute("d")&&(l=`path(${t.getAttribute("d")})`),s.setProperty(i,l,n.getPropertyPriority(i))})}function ur(e,t){A(e,HTMLTextAreaElement)&&(t.innerHTML=e.value),A(e,HTMLInputElement)&&t.setAttribute("value",e.value)}function mr(e,t){if(A(e,HTMLSelectElement)){const r=t,s=Array.from(r.children).find(n=>e.value===n.getAttribute("value"));s&&s.setAttribute("selected","")}}function hr(e,t,r){return A(t,Element)&&(dr(e,t,r),Jt(e,t,r),ur(e,t),mr(e,t)),t}async function fr(e,t){const r=e.querySelectorAll?e.querySelectorAll("use"):[];if(r.length===0)return e;const s={};for(let i=0;i<r.length;i++){const c=r[i].getAttribute("xlink:href");if(c){const d=e.querySelector(c),f=document.querySelector(c);!d&&f&&!s[c]&&(s[c]=await me(f,t,!0))}}const n=Object.values(s);if(n.length){const i="http://www.w3.org/1999/xhtml",l=document.createElementNS(i,"svg");l.setAttribute("xmlns",i),l.style.position="absolute",l.style.width="0",l.style.height="0",l.style.overflow="hidden",l.style.display="none";const c=document.createElementNS(i,"defs");l.appendChild(c);for(let d=0;d<n.length;d++)c.appendChild(n[d]);e.appendChild(l)}return e}async function me(e,t,r){return!r&&t.filter&&!t.filter(e)?null:Promise.resolve(e).then(s=>lr(s,t)).then(s=>cr(e,s,t)).then(s=>hr(e,s,t)).then(s=>fr(s,t))}const et=/url\((['"]?)([^'"]+?)\1\)/g,pr=/url\([^)]+\)\s*format\((["']?)([^"']+)\1\)/g,gr=/src:\s*(?:url\([^)]+\)\s*format\([^)]+\)[,;]\s*)+/g;function xr(e){const t=e.replace(/([.*+?^${}()|\[\]\/\\])/g,"\\$1");return new RegExp(`(url\\(['"]?)(${t})(['"]?\\))`,"g")}function br(e){const t=[];return e.replace(et,(r,s,n)=>(t.push(n),r)),t.filter(r=>!xe(r))}async function yr(e,t,r,s,n){try{const i=r?Ht(t,r):t,l=ye(t);let c;return n||(c=await we(i,l,s)),e.replace(xr(t),`$1${c}$3`)}catch{}return e}function wr(e,{preferredFontFormat:t}){return t?e.replace(gr,r=>{for(;;){const[s,,n]=pr.exec(r)||[];if(!n)return"";if(n===t)return`src: ${s};`}}):e}function tt(e){return e.search(et)!==-1}async function rt(e,t,r){if(!tt(e))return e;const s=wr(e,r);return br(s).reduce((i,l)=>i.then(c=>yr(c,l,t,r)),Promise.resolve(s))}async function se(e,t,r){var s;const n=(s=t.style)===null||s===void 0?void 0:s.getPropertyValue(e);if(n){const i=await rt(n,null,r);return t.style.setProperty(e,i,t.style.getPropertyPriority(e)),!0}return!1}async function vr(e,t){await se("background",e,t)||await se("background-image",e,t),await se("mask",e,t)||await se("-webkit-mask",e,t)||await se("mask-image",e,t)||await se("-webkit-mask-image",e,t)}async function jr(e,t){const r=A(e,HTMLImageElement);if(!(r&&!xe(e.src))&&!(A(e,SVGImageElement)&&!xe(e.href.baseVal)))return;const s=r?e.src:e.href.baseVal,n=await we(s,ye(s),t);await new Promise((i,l)=>{e.onload=i,e.onerror=t.onImageErrorHandler?(...d)=>{try{i(t.onImageErrorHandler(...d))}catch(f){l(f)}}:l;const c=e;c.decode&&(c.decode=i),c.loading==="lazy"&&(c.loading="eager"),r?(e.srcset="",e.src=n):e.href.baseVal=n})}async function _r(e,t){const s=Z(e.childNodes).map(n=>at(n,t));await Promise.all(s).then(()=>e)}async function at(e,t){A(e,Element)&&(await vr(e,t),await jr(e,t),await _r(e,t))}function Sr(e,t){const{style:r}=e;t.backgroundColor&&(r.backgroundColor=t.backgroundColor),t.width&&(r.width=`${t.width}px`),t.height&&(r.height=`${t.height}px`);const s=t.style;return s!=null&&Object.keys(s).forEach(n=>{r[n]=s[n]}),e}const We={};async function Oe(e){let t=We[e];if(t!=null)return t;const s=await(await fetch(e)).text();return t={url:e,cssText:s},We[e]=t,t}async function Ue(e,t){let r=e.cssText;const s=/url\(["']?([^"')]+)["']?\)/g,i=(r.match(/url\([^)]+\)/g)||[]).map(async l=>{let c=l.replace(s,"$1");return c.startsWith("https://")||(c=new URL(c,e.url).href),Je(c,t.fetchRequestInit,({result:d})=>(r=r.replace(l,`url(${d})`),[l,d]))});return Promise.all(i).then(()=>r)}function qe(e){if(e==null)return[];const t=[],r=/(\/\*[\s\S]*?\*\/)/gi;let s=e.replace(r,"");const n=new RegExp("((@.*?keyframes [\\s\\S]*?){([\\s\\S]*?}\\s*?)})","gi");for(;;){const d=n.exec(s);if(d===null)break;t.push(d[0])}s=s.replace(n,"");const i=/@import[\s\S]*?url\([^)]*\)[\s\S]*?;/gi,l="((\\s*?(?:\\/\\*[\\s\\S]*?\\*\\/)?\\s*?@media[\\s\\S]*?){([\\s\\S]*?)}\\s*?})|(([\\s\\S]*?){([\\s\\S]*?)})",c=new RegExp(l,"gi");for(;;){let d=i.exec(s);if(d===null){if(d=c.exec(s),d===null)break;i.lastIndex=c.lastIndex}else c.lastIndex=i.lastIndex;t.push(d[0])}return t}async function kr(e,t){const r=[],s=[];return e.forEach(n=>{if("cssRules"in n)try{Z(n.cssRules||[]).forEach((i,l)=>{if(i.type===CSSRule.IMPORT_RULE){let c=l+1;const d=i.href,f=Oe(d).then(x=>Ue(x,t)).then(x=>qe(x).forEach(u=>{try{n.insertRule(u,u.startsWith("@import")?c+=1:n.cssRules.length)}catch(h){console.error("Error inserting rule from remote css",{rule:u,error:h})}})).catch(x=>{console.error("Error loading remote css",x.toString())});s.push(f)}})}catch(i){const l=e.find(c=>c.href==null)||document.styleSheets[0];n.href!=null&&s.push(Oe(n.href).then(c=>Ue(c,t)).then(c=>qe(c).forEach(d=>{l.insertRule(d,l.cssRules.length)})).catch(c=>{console.error("Error loading remote stylesheet",c)})),console.error("Error inlining remote css file",i)}}),Promise.all(s).then(()=>(e.forEach(n=>{if("cssRules"in n)try{Z(n.cssRules||[]).forEach(i=>{r.push(i)})}catch(i){console.error(`Error while reading CSS rules from ${n.href}`,i)}}),r))}function Cr(e){return e.filter(t=>t.type===CSSRule.FONT_FACE_RULE).filter(t=>tt(t.style.getPropertyValue("src")))}async function Er(e,t){if(e.ownerDocument==null)throw new Error("Provided element is not within a Document");const r=Z(e.ownerDocument.styleSheets),s=await kr(r,t);return Cr(s)}function st(e){return e.trim().replace(/["']/g,"")}function Nr(e){const t=new Set;function r(s){(s.style.fontFamily||getComputedStyle(s).fontFamily).split(",").forEach(i=>{t.add(st(i))}),Array.from(s.children).forEach(i=>{i instanceof HTMLElement&&r(i)})}return r(e),t}async function Rr(e,t){const r=await Er(e,t),s=Nr(e);return(await Promise.all(r.filter(i=>s.has(st(i.style.fontFamily))).map(i=>{const l=i.parentStyleSheet?i.parentStyleSheet.href:null;return rt(i.cssText,l,t)}))).join(`
`)}async function Pr(e,t){const r=t.fontEmbedCSS!=null?t.fontEmbedCSS:t.skipFonts?null:await Rr(e,t);if(r){const s=document.createElement("style"),n=document.createTextNode(r);s.appendChild(n),e.firstChild?e.insertBefore(s,e.firstChild):e.appendChild(s)}}async function $r(e,t={}){const{width:r,height:s}=Ye(e,t),n=await me(e,t,!0);return await Pr(n,t),await at(n,t),Sr(n,t),await Gt(n,r,s)}async function Mr(e,t={}){const{width:r,height:s}=Ye(e,t),n=await $r(e,t),i=await ue(n),l=document.createElement("canvas"),c=l.getContext("2d"),d=t.pixelRatio||Ut(),f=t.canvasWidth||r,x=t.canvasHeight||s;return l.width=f*d,l.height=x*d,t.skipAutoScale||qt(l),l.style.width=`${f}`,l.style.height=`${x}`,t.backgroundColor&&(c.fillStyle=t.backgroundColor,c.fillRect(0,0,l.width,l.height)),c.drawImage(i,0,0,l.width,l.height),l}async function Tr(e,t={}){const r=await Mr(e,t);return await zt(r)}const be="#ffffff";function Fr(){return typeof window>"u"?be:window.getComputedStyle(document.documentElement).getPropertyValue("--bg-elevated").trim()||be}function Lr(e){if(typeof window>"u")return be;const t=window.getComputedStyle(e).backgroundColor;return!t||t==="rgba(0, 0, 0, 0)"||t==="transparent"?Fr():t}async function Ir(e){var r,s;if(typeof window>"u"||!((s=(r=window.navigator)==null?void 0:r.clipboard)!=null&&s.write))throw new Error("Clipboard image copy is not supported in this browser.");if(typeof window.ClipboardItem>"u")throw new Error("Clipboard image copy is not supported in this browser.");const t=await Tr(e,{pixelRatio:2,cacheBust:!0,backgroundColor:Lr(e)});if(!t)throw new Error("Failed to render screenshot.");await window.navigator.clipboard.write([new window.ClipboardItem({"image/png":t})])}const nt=({targetRef:e,label:t})=>{const{toast:r}=ct(),[s,n]=F.useState("idle"),i=F.useRef(null);F.useEffect(()=>()=>{i.current!==null&&window.clearTimeout(i.current)},[]);const l=F.useCallback(async()=>{const d=e.current;if(d){n("copying"),d.classList.add("is-capturing"),d.offsetHeight,await new Promise(f=>window.requestAnimationFrame(()=>f()));try{await Ir(d),n("copied"),r({title:"Screenshot copied",description:`Image of the ${t} copied to your clipboard.`}),i.current!==null&&window.clearTimeout(i.current),i.current=window.setTimeout(()=>n("idle"),1500)}catch(f){n("idle"),r({title:"Copy failed",description:f instanceof Error&&f.message?f.message:"Could not copy the screenshot to your clipboard.",variant:"destructive"})}finally{d.classList.remove("is-capturing")}}},[e,t,r]),c=s==="copied"?"Copied":s==="copying"?"Copying…":"Copy image";return a.jsxs("button",{type:"button",className:"bm-copy-btn",onClick:l,disabled:s==="copying","aria-label":`Copy ${t} screenshot to clipboard`,title:`Copy ${t} screenshot to clipboard`,children:[a.jsx(xt,{name:"copy",size:11}),c]})};function Ar({model:e,datasetName:t,hideClosed:r,setHideClosed:s,setMetric:n}){const i=F.useRef(null),{isInitialFilteredViewLoading:l,chartPoints:c,activeMetric:d,allowedMetrics:f,orderedMetrics:x,parameterBound:u}=e;return l||c.length===0?null:a.jsxs("div",{className:"bm-card",ref:i,children:[a.jsxs("div",{className:"bm-card-screenshot-header","aria-hidden":"true",children:[a.jsxs("div",{className:"bm-card-screenshot-title",children:[a.jsx("span",{className:"bm-card-screenshot-eyebrow",children:"Benchmark"}),a.jsx("span",{className:"bm-card-screenshot-name",children:t})]}),a.jsxs("div",{className:"bm-card-screenshot-brand",children:[a.jsxs("span",{className:"brand",children:[a.jsx("span",{className:"brand-mark"}),a.jsxs("span",{className:"brand-name",children:["Papers ",a.jsx("em",{children:"With Code"})]})]}),a.jsx("span",{className:"bm-card-screenshot-url",children:"paperswithcode.co"})]})]}),a.jsxs("div",{className:"bm-card-head",children:[a.jsx("div",{className:"bm-card-title",children:"SOTA progression"}),a.jsxs("div",{className:"bm-card-actions",children:[f.length>1&&a.jsxs("label",{className:"bm-view-select",children:[a.jsx("span",{className:"bm-view-select-label",children:"View"}),a.jsx("select",{value:d,onChange:h=>n(h.target.value),children:x.map(h=>a.jsx("option",{value:h,children:h},h))})]}),a.jsxs("label",{className:"bm-closed-toggle",title:"Show or hide evaluations of closed-source models",children:[a.jsx(Ke,{className:"bm-closed-toggle-switch",checked:!r,onCheckedChange:h=>{s(!h)},"aria-label":"Show closed-source models"}),a.jsx("span",{className:"bm-closed-toggle-label",children:"Closed-source"})]}),a.jsx(nt,{targetRef:i,label:"SOTA chart"})]})]}),a.jsx("div",{className:"bm-chart-wrap",children:a.jsx(Dt,{points:c,metric:d})}),a.jsxs("div",{className:"bm-chart-foot",children:[a.jsx("span",{className:"legend-dot"}),u===void 0?"All models":"Filtered models"," over time · highlighted dots track the running best · hover for the model · click to open the paper"]})]})}function Br(e){if(!e)return null;const t=e.trim(),r=t.match(/^https?:\/\/(?:www\.)?arxiv\.org\/(?:abs|pdf)\/([^/?#]+)/i);if(r!=null&&r[1])return r[1].replace(/\.pdf$/i,"");const s=t.match(/^(?:https?:\/\/(?:www\.)?paperswithcode\.co)?\/paper\/([^/?#]+)/i);return(s==null?void 0:s[1])??null}function ze(e,t){return e?a.jsx(oe,{title:e}):t||"—"}const ia=()=>{var ve,je,_e,Se,ke,Ce,Ee,Ne,Re,Pe,$e;const{datasetSlug:e}=dt(),[t,r]=ut(),s=t.get("task")||void 0,n=t.get("eval")||void 0,i=St(t.get("max_parameters_exclusive")),{preferences:l,setHideClosedSourceEvals:c}=mt(),d=F.useRef(null),f=F.useRef(!1),x=F.useRef(null),[u,h]=F.useState(null),R=`${e}:${s||""}`,E=u===R,[L,S]=F.useState(1),g=te({queryKey:["dataset",e,"with-paper",s,l.hide_closed_source_evals?"open":"all"],queryFn:()=>re.getDatasetBySlug(e,{include_paper:!0,task:s,is_open:l.hide_closed_source_evals?!0:void 0}),enabled:!!e,staleTime:10*60*1e3}),k=te({queryKey:["dataset-evals",(ve=g.data)==null?void 0:ve.id],queryFn:()=>gt.byDataset(g.data.id),enabled:!!((je=g.data)!=null&&je.id),staleTime:5*60*1e3}),B=te({queryKey:["dataset-tasks",(_e=g.data)==null?void 0:_e.id],queryFn:()=>re.datasetTasks(g.data.id),enabled:!!((Se=g.data)!=null&&Se.id),staleTime:10*60*1e3}),j=te({queryKey:["metric-directions"],queryFn:()=>re.metricDirections(),staleTime:60*60*1e3}),D=Qe(B.data||[],k.data||[],s,n),W=te({queryKey:["task-benchmark-trends",D,l.hide_closed_source_evals?!0:void 0],queryFn:()=>re.trendingBenchmarks(D,{limit:100,min_recent_papers:0,is_open:l.hide_closed_source_evals?!0:void 0}),enabled:!!D&&!!((ke=g.data)!=null&&ke.id),staleTime:10*60*1e3}),w=F.useMemo(()=>{var b;return(((b=W.data)==null?void 0:b.results)||[]).find(M=>{var T;return M.id===((T=g.data)==null?void 0:T.id)})||null},[W.data,(Ce=g.data)==null?void 0:Ce.id]),q=te({queryKey:["leaderboard-metrics",(Ee=g.data)==null?void 0:Ee.id,D],queryFn:()=>re.leaderboardMetrics(g.data.id,D),enabled:!!((Ne=g.data)!=null&&Ne.id)&&!!D,staleTime:10*60*1e3}),I=Tt({evaluations:k.data||[],tasks:B.data||[],configuredMetrics:q.data||[],directions:j.data,initialSummary:(Re=g.data)==null?void 0:Re.leaderboard_summary,taskSlug:s,highlightEvalId:n,metric:t.get("metric")||void 0,hideClosed:l.hide_closed_source_evals,parameterBound:i,showMoreMetrics:E,fittingMetricCount:L,ready:{evaluations:k.isSuccess,tasks:B.isSuccess,directions:j.isSuccess,metrics:q.isSuccess}}),{primaryTask:N,hiddenClosedEvals:Q,hiddenClosedTopEval:z,isSummaryLoading:ee,displayedMetric:O,displayedSota:$,displayedSotaModel:U,isInitialFilteredViewLoading:K,highlightedEval:_,hasParameterMetadata:X,parameterCoverage:Y,orderedMetrics:o}=I;F.useEffect(()=>{if(f.current||!n||!k.data||k.data.length===0)return;const b=d.current;b&&(f.current=!0,b.scrollIntoView({behavior:"smooth",block:"start"}))},[n,k.data]);const m=te({queryKey:["area",N==null?void 0:N.area_id],queryFn:()=>re.getArea(N.area_id),enabled:!!(N!=null&&N.area_id),staleTime:10*60*1e3});if(g.isLoading)return a.jsxs("div",{className:"page",children:[a.jsx(ce,{className:"h-12 w-1/2 mb-6"}),a.jsx(ce,{className:"h-64 w-full"})]});if(g.error||!g.data)return a.jsx("div",{className:"page",children:a.jsx(bt,{variant:"destructive",children:a.jsx(yt,{children:"Benchmark not found."})})});const p=g.data,y=!!p.supports_harness,P=b=>{r(M=>{const T=new URLSearchParams(M);return b===void 0?T.delete("max_parameters_exclusive"):T.set("max_parameters_exclusive",String(b)),T})},C=b=>{r(M=>{const T=new URLSearchParams(M);return T.set("metric",b),T})},v=p.introducing_paper??void 0,H=(v==null?void 0:v.arxiv_id)||(v==null?void 0:v.id)||Br(p.paper_url),J=(v==null?void 0:v.title)||p.introducing_paper_title||null,G=((Pe=v==null?void 0:v.published)==null?void 0:Pe.slice(0,4))||p.introduced_year||null,it=!!H||!!p.paper_url||!!J,lt=v!=null&&v.arxiv_id||($e=p.paper_url)!=null&&$e.match(/^https?:\/\/(?:www\.)?arxiv\.org\//i)?`arXiv:${H}`:`Paper ${H}`;return a.jsx("div",{className:"page",children:a.jsxs("div",{className:"bm-page",children:[a.jsxs("div",{className:"breadcrumb",children:[a.jsx(ne,{to:"/tasks",children:"Tasks"}),m.data&&a.jsxs(a.Fragment,{children:[a.jsx("span",{className:"sep",children:"/"}),a.jsx(ne,{to:`/tasks#area-${jt(m.data.name)}`,children:m.data.name})]}),N&&a.jsxs(a.Fragment,{children:[a.jsx("span",{className:"sep",children:"/"}),a.jsx(ne,{to:`/tasks/${N.slug||N.id}`,children:N.name})]}),a.jsx("span",{className:"sep",children:"/"}),a.jsx("span",{style:{color:"var(--accent-text)"},children:p.name})]}),a.jsxs("div",{className:"bm-title-row",children:[a.jsx("h1",{className:"bm-title",children:p.name}),w&&a.jsxs("div",{className:"bm-title-trend","aria-label":`${p.name} reporting trend`,children:[a.jsx(wt,{history:w.history,benchmarkName:w.name}),a.jsxs("span",{className:"bm-title-trend-summary",children:[a.jsx("strong",{children:w.recent_paper_count})," ",w.recent_paper_count===1?"paper":"papers",w.rate_change==null?w.recent_paper_count>0&&a.jsx("span",{className:"benchmark-trend-change is-new",children:"new"}):a.jsxs("span",{className:`benchmark-trend-change ${w.rate_change>=1?"is-up":"is-down"}`,children:[w.rate_change>=1?"↑":"↓"," ",w.rate_change.toFixed(1),"×"]})]})]})]}),a.jsx(ht,{kind:"benchmark",entityId:p.id}),a.jsxs("div",{className:"bm-subtitle",children:[ee?a.jsx(ce,{className:"h-5 w-64 bm-summary-skeleton","aria-label":"Loading benchmark summary"}):a.jsxs("span",{children:[O&&a.jsxs(a.Fragment,{children:[a.jsx("span",{className:"metric-tag",children:O}),a.jsx("span",{style:{color:"var(--hairline-strong)",margin:"0 8px"},children:"·"})]}),U?a.jsxs("span",{children:["Best: ",a.jsx("span",{className:"best-name",children:le($)})," by ",a.jsx("span",{className:"best-name",style:{fontStyle:"normal"},children:U})]}):a.jsxs("span",{children:[I.rows.length," evaluations"]})]}),p.hf_url&&a.jsxs("a",{className:"bm-action-btn",href:p.hf_url,target:"_blank",rel:"noopener noreferrer",children:[a.jsx("span",{"aria-hidden":"true",children:"🤗"})," View on Hugging Face"]})]}),p.description&&a.jsx("p",{className:"bm-description",children:p.description}),l.hide_closed_source_evals&&Q.length>0&&a.jsxs("div",{className:"bm-hidden-closed-notice",children:[a.jsxs("span",{children:[Q.length," closed-source ",Q.length===1?"row is":"rows are"," hidden",z!=null&&z.model_name?a.jsxs(a.Fragment,{children:[", including ",a.jsx("strong",{children:z.model_name})]}):null,"."]}),a.jsx("button",{type:"button",className:"bm-hidden-closed-action",onClick:()=>{c(!1)},children:"Show closed-source"})]}),it&&a.jsxs("div",{className:"method-source",style:{marginTop:0,marginBottom:18},children:[a.jsx("span",{className:"method-source-label",children:"Source"}),H?a.jsxs(ne,{to:`/paper/${H}`,target:"_blank",rel:"noopener noreferrer",className:"method-source-link",children:[a.jsx("span",{className:"method-source-title",children:J?a.jsx(oe,{title:J}):lt}),G&&a.jsx("span",{className:"method-source-meta",children:G})]}):p.paper_url?a.jsxs("a",{href:vt(p.paper_url)??void 0,target:"_blank",rel:"noopener noreferrer",className:"method-source-link",children:[a.jsx("span",{className:"method-source-title",children:J?a.jsx(oe,{title:J}):p.paper_url}),G&&a.jsx("span",{className:"method-source-meta",children:G})]}):a.jsxs("div",{className:"method-source-link is-static",children:[a.jsx("span",{className:"method-source-title",children:J?a.jsx(oe,{title:J}):null}),G&&a.jsx("span",{className:"method-source-meta",children:G})]})]}),a.jsx(Ar,{model:I,datasetName:p.name,hideClosed:l.hide_closed_source_evals,setHideClosed:c,setMetric:C}),_&&(()=>{var Te,Fe;const b=_.paper_arxiv_id||_.paper_id,M=_.paper_title||(_.paper_arxiv_id?`arXiv:${_.paper_arxiv_id}`:null),T=_.best_metric&&((Te=_.metrics)==null?void 0:Te[_.best_metric])!==void 0?_.best_metric:I.activeMetric,Me=(Fe=_.metrics)==null?void 0:Fe[T];return a.jsxs("div",{className:"bm-highlight-banner",role:"status",children:[a.jsx("span",{className:"bm-highlight-dot","aria-hidden":"true"}),a.jsxs("div",{className:"bm-highlight-text",children:[a.jsx("span",{className:"bm-highlight-label",children:"Comparing result"}),a.jsxs("span",{className:"bm-highlight-paper",children:[_.model_name||M||"—",_.harness?` · ${_.harness}`:""]}),a.jsxs("span",{className:"bm-highlight-meta",children:[Me!==void 0&&T?a.jsxs(a.Fragment,{children:[T,": ",le(Me),M?" · ":""]}):null,b?a.jsx(ne,{to:`/paper/${b}`,children:ze(_.paper_title,M)}):ze(_.paper_title,M)]})]})]})})(),a.jsxs("div",{className:"bm-card",ref:x,children:[a.jsxs("div",{className:"bm-card-screenshot-header","aria-hidden":"true",children:[a.jsxs("div",{className:"bm-card-screenshot-title",children:[a.jsx("span",{className:"bm-card-screenshot-eyebrow",children:"Benchmark"}),a.jsx("span",{className:"bm-card-screenshot-name",children:p.name})]}),a.jsxs("div",{className:"bm-card-screenshot-brand",children:[a.jsxs("span",{className:"brand",children:[a.jsx("span",{className:"brand-mark"}),a.jsxs("span",{className:"brand-name",children:["Papers ",a.jsx("em",{children:"With Code"})]})]}),a.jsx("span",{className:"bm-card-screenshot-url",children:"paperswithcode.co"})]})]}),a.jsxs("div",{className:"bm-card-head",children:[a.jsx("div",{className:"bm-card-title",children:"Leaderboard"}),a.jsxs("div",{className:"bm-card-actions",children:[I.allowedMetrics.length>1&&a.jsxs("label",{className:"bm-view-select",children:[a.jsx("span",{className:"bm-view-select-label",children:"View"}),a.jsx("select",{value:I.activeMetric,onChange:b=>C(b.target.value),children:o.map(b=>a.jsx("option",{value:b,children:b},b))})]}),a.jsxs("label",{className:"bm-closed-toggle",title:"Show or hide evaluations of closed-source models",children:[a.jsx(Ke,{className:"bm-closed-toggle-switch",checked:!l.hide_closed_source_evals,onCheckedChange:b=>{c(!b)},"aria-label":"Show closed-source models"}),a.jsx("span",{className:"bm-closed-toggle-label",children:"Closed-source"})]}),!K&&!X&&k.isSuccess?a.jsx("span",{className:"bm-parameter-unavailable",children:"Parameter metadata unavailable"}):null,a.jsx(nt,{targetRef:x,label:"leaderboard"})]})]}),!K&&X&&a.jsxs("div",{className:"bm-parameter-filter-bar",children:[a.jsx("span",{className:"bm-parameter-filter-label",children:"Parameters Size"}),a.jsx("div",{className:"bm-parameter-filter-options",role:"group","aria-label":"Filter leaderboard by parameter size",children:I.parameterFilters.map(b=>{const{selected:M,disabled:T}=b;return a.jsx("button",{type:"button",className:`bm-parameter-filter-option${M?" is-active":""}`,"aria-pressed":M,disabled:T,onClick:()=>P(b.maxParametersExclusive),children:b.label},b.label)})}),a.jsxs("span",{className:"bm-parameter-coverage",children:["Parameters known for ",Y.parameter_coverage_known," of ",Y.parameter_coverage_total," entries"]})]}),k.isLoading||K?a.jsx("div",{style:{padding:16},children:Array.from({length:5}).map((b,M)=>a.jsx(ce,{className:"h-10 w-full mb-2"},M))}):I.rows.length===0?i?a.jsxs("div",{className:"no-results",children:["No known parameter counts match this filter. "," ",a.jsx("button",{type:"button",className:"bm-empty-filter-action",onClick:()=>P(void 0),children:"Show all"})]}):a.jsx("div",{className:"no-results",children:"No evaluations submitted yet."}):a.jsxs(a.Fragment,{children:[o.length>L&&a.jsxs("div",{className:"bm-metric-columns-control",children:[a.jsx("button",{type:"button",className:"bm-metric-columns-toggle","aria-expanded":E,"aria-controls":"benchmark-leaderboard",onClick:()=>h(E?null:R),children:E?"Show fewer metrics":`Show more metrics (${o.length-L})`}),E&&a.jsx("span",{children:"Scroll metrics horizontally; model and paper details stay visible."})]}),a.jsx(At,{model:I,showHarness:y,setMetric:C,highlightRowRef:d,onFittingMetricCountChange:S})]})]})]})})};export{ia as default};