openpond 0.2.61 → 0.2.63

package/dist/web/assets/LabModelVersionDetailPage-CF8FkRUY.js

@@ -1 +0,0 @@
-import{i as e}from"./rolldown-runtime-Dd_uD5pT.js";import{g as t,m as n}from"./index-BsPUniZy.js";import"./icons-kaq6k88m.js";import{c as r,l as i,n as a,o,r as s,t as c}from"./training-model-data-CmVo80ZR.js";import{t as l}from"./chevron-down-Bnma8EMP.js";import{t as u}from"./chevron-right-Cxv_xBz5.js";import{Mn as d,Nn as f,_n as p,qn as m,rr as h}from"./App-Yy3l2sdD.js";import{l as g}from"./MainPane-BpE7AKXu.js";import{C as _,S as v,_ as y,b,c as x,d as S,f as ee,g as C,h as w,l as te,m as ne,p as re,s as ie,u as ae,v as oe,x as T,y as E}from"./LabsRoute-DOnw39du.js";var D=e(t(),1),O=n();function k({events:e}){let t=e.filter(e=>e.type===`metric`&&e.payload.metricKind===`rollout_trajectory`);return t.length?(0,O.jsx)(`div`,{className:`training-rollout-receipts`,children:t.map((e,t)=>{let n=e.payload.rewardEligible===!0,r=j(e.payload.reward),i=typeof e.payload.failureClass==`string`?e.payload.failureClass:null,a=typeof e.payload.failureCode==`string`?e.payload.failureCode:null;return(0,O.jsxs)(`details`,{children:[(0,O.jsxs)(`summary`,{children:[(0,O.jsxs)(`span`,{children:[`Attempt `,t+1]}),(0,O.jsx)(`strong`,{children:n?`eligible`:`ineligible`}),(0,O.jsx)(`em`,{children:n&&r!=null?r.toFixed(3):a??i??`not reward eligible`})]}),(0,O.jsxs)(`dl`,{className:`training-configuration-list`,children:[(0,O.jsx)(A,{label:`Rollout`,value:M(e.payload.rolloutId)}),(0,O.jsx)(A,{label:`Group`,value:M(e.payload.rolloutGroupId)}),(0,O.jsx)(A,{label:`Policy`,value:`v${j(e.payload.policyVersion)??0}`}),(0,O.jsx)(A,{label:`Worker`,value:String((j(e.payload.workerSlot)??0)+1)}),(0,O.jsx)(A,{label:`Reward`,value:r==null?`Not eligible`:String(r)}),(0,O.jsx)(A,{label:`Failure class`,value:i??`None`}),(0,O.jsx)(A,{label:`Failure code`,value:a??`None`}),(0,O.jsx)(A,{label:`Components`,value:JSON.stringify(e.payload.rewardComponents??{})}),(0,O.jsx)(A,{label:`Tokens`,value:`${j(e.payload.inputTokens)??0} input · ${j(e.payload.outputTokens)??0} output`})]})]},e.id)})}):(0,O.jsx)(`div`,{className:`training-run-placeholder`,children:`No managed attempts were recorded for this run.`})}function se({receipts:e}){let[t,n]=(0,D.useState)(!1);if(!e.length)return(0,O.jsx)(`div`,{className:`training-run-placeholder`,children:`No on-policy rollout receipts were recorded for this run.`});let r=[...e].sort((e,t)=>t.updatedAt.localeCompare(e.updatedAt)),i=t?r:r.slice(0,24);return(0,O.jsxs)(`div`,{className:`training-rollout-receipts`,children:[e.length>24?(0,O.jsxs)(`div`,{className:`training-rollout-receipt-controls`,children:[(0,O.jsx)(`span`,{children:t?`Showing all ${e.length} receipts`:`Showing latest ${i.length} of ${e.length} receipts`}),(0,O.jsx)(`button`,{className:`training-text-button`,type:`button`,onClick:()=>n(e=>!e),children:t?`Show latest 24`:`Show all ${e.length}`})]}):null,i.map(e=>(0,O.jsxs)(`details`,{children:[(0,O.jsxs)(`summary`,{children:[(0,O.jsx)(`span`,{children:ce(e.providerTrace.rolloutId)}),(0,O.jsx)(`strong`,{children:e.status}),(0,O.jsx)(`em`,{children:e.reward.eligible&&e.reward.raw!=null?e.reward.raw.toFixed(3):e.failureClass??`not reward eligible`})]}),(0,O.jsxs)(`dl`,{className:`training-configuration-list`,children:[(0,O.jsx)(A,{label:`Correlation`,value:e.correlationId}),(0,O.jsx)(A,{label:`Task`,value:e.taskId}),(0,O.jsx)(A,{label:`Policy`,value:e.policy.modelId}),(0,O.jsx)(A,{label:`Checkpoint`,value:e.policy.checkpointId??`Provider initial policy`}),(0,O.jsx)(A,{label:`Environment`,value:`${e.environment.id}@${e.environment.version}`}),(0,O.jsx)(A,{label:`World`,value:`${e.environment.worldId} · ${e.environment.worldHash}`}),(0,O.jsx)(A,{label:`Raw reward`,value:e.reward.raw==null?`Not eligible`:String(e.reward.raw)}),(0,O.jsx)(A,{label:`Components`,value:JSON.stringify(e.reward.components)}),(0,O.jsx)(A,{label:`Outcome`,value:e.verifier?.outcome??e.failureClass??`Pending`}),(0,O.jsx)(A,{label:`Trace`,value:`${e.providerTrace.experimentId}/${e.providerTrace.rolloutId}`})]})]},e.id))]})}function A({label:e,value:t}){return(0,O.jsxs)(`div`,{children:[(0,O.jsx)(`dt`,{children:e}),(0,O.jsx)(`dd`,{children:t})]})}function j(e){return typeof e==`number`&&Number.isFinite(e)?e:null}function M(e){return typeof e==`string`&&e?e:`Not recorded`}function ce(e){return e.replace(/^lineage_/,``).slice(0,18)}function le({detail:e,loading:t,pending:n=!1}){if(t&&!e)return(0,O.jsx)(`div`,{className:`training-run-placeholder`,children:`Loading evaluation results…`});let r=e?.evaluation;return r?(0,O.jsxs)(`div`,{className:`training-run-evaluation`,children:[(0,O.jsxs)(`div`,{className:`training-metric-summary`,children:[(0,O.jsx)(P,{label:`Base score`,value:ue(r.base.meanScore)}),(0,O.jsx)(P,{label:`Trained score`,value:ue(r.trained.meanScore)}),(0,O.jsx)(P,{label:`Score change`,value:fe(r.meanScoreDelta)}),(0,O.jsx)(P,{label:`Trained pass rate`,value:de(r.trained.passRate)})]}),r.base.count===0?(0,O.jsx)(`p`,{className:`training-muted`,children:`Base-model scoring was not imported for this earlier run. New runs retain both base and trained evaluations.`}):null,r.trained.count>0&&r.trained.scoredCount===0?(0,O.jsx)(`p`,{className:`training-muted`,children:`Outputs were captured, but the grader was unavailable, so this run has no evaluation score.`}):null,(0,O.jsx)(`div`,{className:`training-table-wrap`,children:(0,O.jsxs)(`table`,{className:`training-data-table training-evaluation-table`,children:[(0,O.jsx)(`thead`,{children:(0,O.jsxs)(`tr`,{children:[(0,O.jsx)(`th`,{children:`Test example`}),(0,O.jsx)(`th`,{children:`Base`}),(0,O.jsx)(`th`,{children:`Trained`}),(0,O.jsx)(`th`,{children:`Result`})]})}),(0,O.jsx)(`tbody`,{children:r.examples.map(e=>(0,O.jsxs)(`tr`,{children:[(0,O.jsx)(`td`,{children:(0,O.jsxs)(`details`,{children:[(0,O.jsx)(`summary`,{children:e.taskId}),(0,O.jsx)(N,{input:e.input,baseOutput:e.baseOutput,trainedOutput:e.trainedOutput,baseGrade:e.baseGrade,trainedGrade:e.trainedGrade})]})}),(0,O.jsx)(`td`,{children:pe(e.baseGrade)}),(0,O.jsx)(`td`,{children:pe(e.trainedGrade)}),(0,O.jsx)(`td`,{children:F(e.baseGrade,e.trainedGrade)})]},e.taskId))})]})})]}):(0,O.jsx)(`div`,{className:`training-run-placeholder`,children:n?`Evaluation will appear here after the first eligible checkpoint is scored.`:`No scored evaluation results are available for this run.`})}function N({input:e,baseOutput:t,trainedOutput:n,baseGrade:r,trainedGrade:i}){let a=[...new Set([...r?.feedback??[],...i?.feedback??[]])];return(0,O.jsxs)(`div`,{className:`training-evaluation-output`,children:[(0,O.jsxs)(`div`,{children:[(0,O.jsx)(`span`,{children:`Input`}),(0,O.jsx)(`pre`,{children:JSON.stringify(e,null,2)})]}),t?(0,O.jsxs)(`div`,{children:[(0,O.jsx)(`span`,{children:`Base output`}),(0,O.jsx)(`pre`,{children:JSON.stringify(t,null,2)})]}):null,n?(0,O.jsxs)(`div`,{children:[(0,O.jsx)(`span`,{children:`Trained output`}),(0,O.jsx)(`pre`,{children:JSON.stringify(n,null,2)})]}):null,a.length?(0,O.jsxs)(`div`,{children:[(0,O.jsx)(`span`,{children:`Grader feedback`}),(0,O.jsx)(`p`,{children:a.join(` `)})]}):null]})}function P({label:e,value:t}){return(0,O.jsxs)(`div`,{children:[(0,O.jsx)(`span`,{children:e}),(0,O.jsx)(`strong`,{children:t})]})}function ue(e){return e==null?`Not recorded`:e.toFixed(3)}function de(e){return e==null?`Not recorded`:`${Math.round(e*100)}%`}function fe(e){return e==null?`Not comparable`:`${e>=0?`+`:``}${e.toFixed(3)}`}function pe(e){return e?.status===`unavailable`?`Unavailable`:e?.score==null?`Not scored`:e.score.toFixed(3)}function F(e,t){if(!t)return`Not evaluated`;if(t.status===`unavailable`)return`Unavailable`;if(!e||e.score==null||t.score==null)return t.passed?`Passed`:`Failed`;let n=t.score-e.score;return Math.abs(n)<5e-4?`No change`:n>0?`Improved ${n.toFixed(3)}`:`Regressed ${Math.abs(n).toFixed(3)}`}function me({connection:e,lineageId:t,jobId:n,onSaved:r}){let i=n??t,a=n?`jobs`:`models`,[o,s]=(0,D.useState)(null),[c,l]=(0,D.useState)(0),[u,d]=(0,D.useState)(!0),[f,m]=(0,D.useState)(!1),[g,_]=(0,D.useState)(!1),[v,y]=(0,D.useState)(null),[b,x]=(0,D.useState)(null),[S,ee]=(0,D.useState)(``),[C,w]=(0,D.useState)(null),[te,ne]=(0,D.useState)([]),[re,ie]=(0,D.useState)(!1);(0,D.useEffect)(()=>{if(!e){d(!1),m(!1),y(`Connect to OpenPond to review this candidate.`);return}let t=!0;return d(!0),m(!1),y(null),ne([]),(async()=>{try{if(a===`models`){let n=await h.candidateReview(e,i);t&&n&&s(n)}let n=await h.candidateReview(e,i,!0,a);t&&(s(n),m(!!n))}catch(e){t&&y(I(e))}finally{t&&d(!1)}})(),()=>{t=!1}},[e,i,a,c]);function ae(e){x(e),ee(``),w(null)}async function oe(){if(!e||!o||!b||!S.trim())return;let t=C&&C.reason===S.trim()&&C.decision===b?C:{schemaVersion:`openpond.trainingCandidateDecisionRequest.v1`,...o.target,decision:b,reason:S.trim(),idempotencyKey:crypto.randomUUID(),expectedDecision:o.decision?{id:o.decision.id,contentHash:o.decision.contentHash}:null};w(t),_(!0),y(null);try{s(await h.recordCandidateReview(e,i,t,a)),x(null),w(null),ne([]),await r()}catch(e){y(I(e))}finally{_(!1)}}let T=(te.at(-1)??o?.decision)?.request.expectedDecision;async function E(){if(e&&T){ie(!0),y(null);try{let t=await h.candidateReviewHistory(e,i,T,a);ne(e=>[...e,t])}catch(e){y(I(e))}finally{ie(!1)}}}let k=o?.decision;return(0,O.jsxs)(`section`,{className:`training-run-evaluation`,"aria-label":`Candidate review`,children:[(0,O.jsx)(`h3`,{children:`Candidate review`}),k?(0,O.jsxs)(O.Fragment,{children:[(0,O.jsx)(`strong`,{children:k.request.decision===`accepted`?`Accepted`:`Rejected`}),(0,O.jsx)(`p`,{children:k.request.reason}),(0,O.jsxs)(`p`,{className:`training-muted`,children:[`Review `,k.revision,` · `,new Date(k.decidedAt).toLocaleString()]})]}):(0,O.jsx)(`p`,{className:`training-muted`,children:u?`Loading candidate review…`:o?`No candidate decision recorded.`:`Candidate review unavailable.`}),o?(0,O.jsxs)(`p`,{className:`training-muted`,children:[`Last synced `,new Date(o.syncedAt).toLocaleString(),`. Acceptance keeps this candidate for later use; serving is activated separately.`]}):null,v?(0,O.jsx)(p,{error:v}):null,(0,O.jsxs)(`div`,{className:`training-dialog-actions`,children:[(0,O.jsx)(`button`,{className:`training-button secondary`,type:`button`,disabled:u||g,onClick:()=>l(e=>e+1),children:`Refresh review`}),(0,O.jsx)(`button`,{className:`training-button`,type:`button`,disabled:!f||u||g||k?.request.decision===`accepted`,onClick:()=>ae(`accepted`),children:`Accept candidate`}),(0,O.jsx)(`button`,{className:`training-button danger`,type:`button`,disabled:!f||u||g||k?.request.decision===`rejected`,onClick:()=>ae(`rejected`),children:`Reject candidate`})]}),te.map(e=>(0,O.jsxs)(`details`,{className:`training-evidence`,children:[(0,O.jsxs)(`summary`,{children:[`Review `,e.revision,` · `,e.request.decision]}),(0,O.jsx)(`p`,{children:e.request.reason}),(0,O.jsx)(`p`,{children:new Date(e.decidedAt).toLocaleString()})]},e.id)),T?(0,O.jsx)(`button`,{className:`training-button secondary`,type:`button`,disabled:re||g,onClick:()=>void E(),children:re?`Loading…`:`Earlier review`}):null,b?(0,O.jsx)(`div`,{className:`training-dialog-backdrop`,role:`presentation`,onMouseDown:()=>!g&&x(null),children:(0,O.jsxs)(`section`,{className:`training-dialog training-promotion-dialog`,role:`dialog`,"aria-modal":`true`,"aria-label":`Review candidate`,onMouseDown:e=>e.stopPropagation(),children:[(0,O.jsx)(`div`,{className:`training-dialog-header`,children:(0,O.jsx)(`h2`,{children:b===`accepted`?`Accept this candidate?`:`Reject this candidate?`})}),(0,O.jsx)(`p`,{children:`The adapter and its recorded evaluation remain in the candidate’s history.`}),(0,O.jsxs)(`label`,{className:`training-promotion-reason`,children:[(0,O.jsx)(`span`,{children:`Reason`}),(0,O.jsx)(`textarea`,{autoFocus:!0,maxLength:5e3,value:S,onChange:e=>ee(e.target.value)})]}),v?(0,O.jsx)(p,{error:v}):null,(0,O.jsxs)(`div`,{className:`training-dialog-actions`,children:[(0,O.jsx)(`button`,{className:`training-button secondary`,type:`button`,disabled:g,onClick:()=>x(null),children:`Cancel`}),(0,O.jsx)(`button`,{className:`training-button ${b===`rejected`?`danger`:``}`,type:`button`,disabled:g||!S.trim(),onClick:()=>void oe(),children:g?`Saving…`:C?`Retry save`:`Save decision`})]})]})}):null]})}function I(e){return e instanceof Error?e.message:`Candidate review could not be loaded.`}function he({connection:e,jobId:t,evaluations:n}){let[r,i]=(0,D.useState)(null),a=n.filter(e=>e.reference),o=a.find(e=>e.reference?.id===r)??a.filter(e=>e.kind===`candidate`).at(-1)??a[0];return!e||!o?.reference?null:(0,O.jsxs)(`section`,{className:`training-run-evaluation`,"aria-label":`Retained evaluation answers`,children:[(0,O.jsxs)(`label`,{className:`training-taskset-selector`,children:[(0,O.jsx)(`span`,{children:`Evaluation answers`}),(0,O.jsx)(`select`,{value:o.reference.id,onChange:e=>i(e.target.value),children:a.map(e=>(0,O.jsxs)(`option`,{value:e.reference.id,children:[e.kind===`baseline`?`Baseline`:`Candidate`,` · policy `,e.policyVersion]},e.reference.id))})]}),(0,O.jsx)(ge,{connection:e,jobId:t,evaluation:o.reference},`${t}:${o.reference.id}:${o.reference.contentHash}`)]})}function ge({connection:e,jobId:t,evaluation:n}){let[r,i]=(0,D.useState)(),[a,o]=(0,D.useState)([]),[s,c]=(0,D.useState)(null),[l,u]=(0,D.useState)(null),[d,f]=(0,D.useState)(!0),[m,g]=(0,D.useState)(0),{id:_,contentHash:v}=n;return(0,D.useEffect)(()=>{let n=!0;return f(!0),u(null),c(null),h.trainingEvaluationTasks(e,t,{id:_,contentHash:v},r).then(e=>{n&&c(e)}).catch(e=>{n&&u(e instanceof Error?e.message:String(e))}).finally(()=>{n&&f(!1)}),()=>{n=!1}},[e,t,_,v,r,m]),(0,O.jsxs)(`div`,{className:`training-run-evaluation`,"aria-busy":d,children:[d?(0,O.jsx)(`p`,{className:`training-muted`,children:`Loading recorded answers…`}):null,l?(0,O.jsx)(p,{error:l,children:(0,O.jsx)(`button`,{className:`training-button secondary`,type:`button`,onClick:()=>g(e=>e+1),children:`Retry answers`})}):null,s?(0,O.jsxs)(O.Fragment,{children:[(0,O.jsxs)(`p`,{className:`training-muted`,children:[`Tasks `,s.offset+1,`–`,s.offset+s.tasks.length,` of`,` `,s.total,`. Recorded outputs and scores are retained for this run.`]}),s.tasks.map((e,t)=>(0,O.jsxs)(`details`,{className:`training-evidence`,children:[(0,O.jsxs)(`summary`,{children:[`Example `,s.offset+t+1,` · score `,e.score.toFixed(3)]}),(0,O.jsxs)(`div`,{className:`training-evaluation-output`,children:[(0,O.jsxs)(`div`,{children:[(0,O.jsx)(`span`,{children:`Instruction`}),(0,O.jsx)(`pre`,{children:e.input.instruction})]}),(0,O.jsxs)(`div`,{children:[(0,O.jsx)(`span`,{children:`Context`}),(0,O.jsx)(`pre`,{children:JSON.stringify(e.input.context,null,2)})]}),(0,O.jsxs)(`div`,{children:[(0,O.jsx)(`span`,{children:`Recorded output`}),(0,O.jsx)(`pre`,{children:e.output})]})]})]},e.taskId))]}):null,(0,O.jsxs)(`div`,{className:`training-dialog-actions`,children:[(0,O.jsx)(`button`,{className:`training-button secondary`,type:`button`,disabled:d||!a.length,onClick:()=>{i(a.at(-1)),o(e=>e.slice(0,-1))},children:`Previous tasks`}),(0,O.jsx)(`button`,{className:`training-button secondary`,type:`button`,disabled:d||!s?.nextCursor,onClick:()=>{s?.nextCursor&&(o(e=>[...e,r]),i(s.nextCursor))},children:`Next tasks`})]})]})}var _e=new Set([`queued`,`starting`,`running`,`cancelling`,`reconciling`]),L=new WeakMap;function ve(e,t,n,r=!1){let[i,a]=(0,D.useState)(null),[o,s]=(0,D.useState)(!1),[c,l]=(0,D.useState)(null),u=(0,D.useRef)(t),d=(0,D.useRef)(null);u.current=t;let f=(0,D.useCallback)((n=r)=>{if(!e||!t)return Promise.resolve(null);let i=`${t}:${n?`full`:`live`}`;if(d.current?.key===i)return d.current.request;s(!0);let o=h.trainingRunDetail(e,t,{includeEvaluation:n}).then(r=>{let i=R(e),o=i.get(t),s=!n&&o?.includesEvaluation?{...r,evaluation:o.detail.evaluation}:r;return i.set(t,{detail:s,includesEvaluation:n||o?.includesEvaluation===!0}),u.current===t&&(a(s),l(null)),s}).catch(e=>(u.current===t&&l(e instanceof Error?e.message:String(e)),null)).finally(()=>{d.current?.request===o&&(d.current=null,s(!1))});return d.current={key:i,request:o},o},[e,r,t]);return(0,D.useEffect)(()=>{if(l(null),!e||!t){a(null);return}let n=R(e).get(t);a(n?.detail??null),f(r)},[e,r,t,n,f]),(0,D.useEffect)(()=>{if(!e||!t||!n||!_e.has(n))return;let r=window.setInterval(()=>void f(!1),2e3);return()=>window.clearInterval(r)},[e,t,n,f]),{detail:i,error:c,loading:o,refresh:f}}function R(e){let t=L.get(e);if(t)return t;let n=new Map;return L.set(e,n),n}function ye({title:e,status:t,statusValue:n,versionStatus:r,method:i,baseModel:a,taskset:o,compute:s,duration:c,output:l,reward:d,telemetry:f,evidence:p,configuration:m,children:h,failure:g,onOpenTaskset:v,showHeader:y=!0}){let b=f?.cost.providerReportedUsd??f?.cost.estimatedUsd??p?.cost.totalUsd??null;return(0,O.jsxs)(`div`,{className:`labs-model-run-summary`,children:[y?(0,O.jsx)(`section`,{className:`labs-run-outcome-card`,children:(0,O.jsx)(`header`,{children:(0,O.jsxs)(`div`,{children:[(0,O.jsxs)(`div`,{className:`labs-run-title-row`,children:[(0,O.jsx)(`h2`,{children:e}),(0,O.jsx)(_,{label:t,value:n})]}),(0,O.jsxs)(`p`,{children:[U(i),` on `,a]})]})})}):null,h,(0,O.jsxs)(`section`,{className:`labs-run-summary-card`,children:[(0,O.jsx)(`header`,{children:(0,O.jsx)(`h3`,{children:`Run setup`})}),(0,O.jsxs)(`dl`,{className:`labs-run-detail-list`,children:[(0,O.jsx)(z,{label:`Training method`,value:i}),(0,O.jsx)(z,{label:`Base model`,value:a}),(0,O.jsx)(z,{label:`Compute`,value:s}),(0,O.jsx)(z,{label:`Output`,value:l}),r?(0,O.jsx)(z,{label:`Version status`,value:r}):null,(0,O.jsx)(z,{label:`Duration`,value:c}),(0,O.jsx)(z,{label:`Run cost`,value:b===null?`Not reported`:`$${b.toFixed(4)}`}),(0,O.jsx)(z,{label:`Final reward`,value:d===null?`Not reported`:d.toFixed(4)}),m.map(e=>(0,O.jsx)(z,{label:e.label,value:e.value},e.label)),f||p?(0,O.jsxs)(O.Fragment,{children:[(0,O.jsx)(B,{label:`GPU`,value:V(f,p)}),(0,O.jsx)(B,{label:`Tokens`,value:`${H(f?.usage.promptTokens??p?.usage.inputTokens??null)} prompt · ${H(f?.usage.generatedTokens??p?.usage.outputTokens??null)} generated`}),(0,O.jsx)(B,{label:`Trajectories`,value:`${H(f?.usage.successfulTrajectories??p?.reward.eligibleTrajectoryCount??null)} succeeded · ${H(f?.usage.failedTrajectories??(p?p.reward.trajectoryCount-p.reward.eligibleTrajectoryCount:null))} failed`}),(0,O.jsx)(B,{label:`Optimizer updates`,value:H(f?.usage.optimizerSteps??p?.progress.committedOptimizerSteps??null)}),(0,O.jsx)(B,{label:`GPU time`,value:f?.usage.gpuSeconds==null?`Not reported`:`${f.usage.gpuSeconds.toFixed(1)} seconds`})]}):(0,O.jsx)(z,{label:`Resource usage`,value:`Not retained for this run`}),(0,O.jsxs)(`div`,{children:[(0,O.jsx)(`dt`,{children:`Taskset`}),(0,O.jsx)(`dd`,{children:v?(0,O.jsxs)(`button`,{className:`labs-run-taskset-link`,type:`button`,onClick:v,children:[(0,O.jsx)(`span`,{children:o}),(0,O.jsx)(u,{"aria-hidden":`true`,size:14})]}):o})]})]})]}),g?(0,O.jsx)(`p`,{className:`labs-training-error`,children:g}):null]})}function z({label:e,value:t}){return(0,O.jsxs)(`div`,{children:[(0,O.jsx)(`dt`,{children:e}),(0,O.jsx)(`dd`,{children:t})]})}function B({label:e,value:t}){return(0,O.jsx)(z,{label:e,value:t})}function V(e,t){let n=e?.resource.gpuCount??t?.resource.gpuCount??null,r=e?.resource.gpuType??t?.resource.gpuType??null;return n===null&&r===null?`Not reported`:`${n??`Unknown`} × ${r?be(r):`Unreported`}`}function H(e){return e===null?`Not reported`:e.toLocaleString()}function U(e){return e===`RFT`?`Reinforcement fine-tuning`:`${e} training`}function be(e){return e.split(` · `).map(e=>e.replaceAll(`_`,` `).replace(/\b[a-z]/g,e=>e.toUpperCase())).join(` · `)}function W({receipt:e,onOpenConversation:t}){return(0,O.jsxs)(O.Fragment,{children:[(0,O.jsxs)(`section`,{className:`labs-run-summary-card`,children:[(0,O.jsx)(`header`,{children:(0,O.jsx)(`h3`,{children:`Inconclusive benchmark`})}),(0,O.jsx)(`p`,{className:`training-run-placeholder`,children:e.reason}),(0,O.jsxs)(`dl`,{className:`labs-run-detail-list`,children:[(0,O.jsx)(q,{label:`Held-out baseline tokens`,value:e.usage.baseline.totalTokens.toLocaleString()}),(0,O.jsx)(q,{label:`Adaptation baseline tokens`,value:e.usage.adaptation.totalTokens.toLocaleString()}),(0,O.jsx)(q,{label:`Refiner tokens`,value:e.usage.refiner.totalTokens.toLocaleString()}),(0,O.jsx)(q,{label:`Grader tokens`,value:e.usage.grader.totalTokens.toLocaleString()}),(0,O.jsx)(q,{label:`Observed / maximum spend`,value:`$${e.budget.observedSpendUsd.toFixed(4)} / $${e.budget.maximumSpendUsd.toFixed(2)}`}),(0,O.jsx)(q,{label:`Harness candidate`,value:e.stopReason===`candidate_harness_unchanged`?`Unchanged`:`Lineage invalid`}),(0,O.jsx)(q,{label:`Completed attempts`,value:e.attempts.length.toLocaleString()}),(0,O.jsx)(q,{label:`Candidate attempts`,value:`Skipped`})]})]}),e.attempts.length?(0,O.jsx)(K,{receipt:e,onOpenConversation:t}):null]})}function xe({run:e}){let t=[`baseline`,`adaptation`,`refiner`,`candidate_adaptation`,`candidate`,`comparison`],n={baseline:`Baseline`,adaptation:`Adaptation`,refiner:`Refiner`,candidate_adaptation:`Sequential adaptation`,candidate:`Candidate`,comparison:`Comparison`},r=Math.max(0,t.indexOf(e.evaluationProgress?.stage??`baseline`)),i=e.evaluationProgress?.completedAttempts??0,a=e.evaluationProgress?.totalAttempts??1;return(0,O.jsxs)(`div`,{className:`labs-benchmark-progress`,children:[(0,O.jsxs)(`header`,{children:[(0,O.jsx)(`h3`,{children:`Benchmark progress`}),(0,O.jsxs)(`strong`,{children:[i,` of `,a,` attempts`]})]}),(0,O.jsx)(`div`,{className:`labs-benchmark-progress-track`,"aria-hidden":`true`,children:(0,O.jsx)(`span`,{style:{width:`${Math.min(100,i/a*100)}%`}})}),(0,O.jsx)(`ol`,{children:t.map((e,t)=>(0,O.jsxs)(`li`,{className:t<r?`complete`:t===r?`active`:void 0,children:[(0,O.jsx)(`span`,{}),n[e]]},e))})]})}function Se({receipt:e}){let t=T(e),n=t.pairs.filter(e=>e.cohort===`held_out`),r=t.pairs.filter(e=>e.cohort===`adaptation`),i=e=>e.length?e.reduce((e,t)=>e+t,0)/e.length:0;return(0,O.jsxs)(`div`,{className:`labs-benchmark-attempt-charts`,children:[(0,O.jsx)(G,{title:`Held-out tokens by task`,baseline:n.map(e=>e.baseline.totalTokens),candidate:n.map(e=>e.refined.totalTokens),format:e=>Math.round(e).toLocaleString()}),(0,O.jsx)(G,{title:`Adaptation tokens by task`,baseline:r.map(e=>e.baseline.totalTokens),candidate:r.map(e=>e.refined.totalTokens),format:e=>Math.round(e).toLocaleString()}),(0,O.jsx)(G,{title:`Held-out quality by task`,baseline:n.map(e=>e.baseline.score),candidate:n.map(e=>e.refined.score),format:e=>e.toFixed(2),summarize:i}),(0,O.jsx)(G,{title:`Adaptation quality by task`,baseline:r.map(e=>e.baseline.score),candidate:r.map(e=>e.refined.score),format:e=>e.toFixed(2),summarize:i})]})}function Ce({receipt:e,run:t,tasksetName:n,onOpenTaskset:r}){let i=T(e),a=i.comparedTaskCount>0,s=i.passed,c=i.tokenDeltaPercent===null?`—`:`${i.tokenDeltaPercent>0?`+`:``}${i.tokenDeltaPercent.toFixed(2)}%`,l=e=>{let t=i.cohorts[e];return`${t.passedTaskCount}/${t.comparedTaskCount} passed`};return(0,O.jsxs)(`section`,{className:`labs-run-summary-card`,children:[(0,O.jsx)(`header`,{children:(0,O.jsx)(`h3`,{children:`Token efficiency`})}),a?(0,O.jsxs)(`div`,{className:`labs-benchmark-efficiency-summary`,children:[(0,O.jsxs)(`strong`,{children:[i.passedTaskCount,` of `,i.comparedTaskCount,` tasks passed`]}),(0,O.jsxs)(`p`,{children:[`A task passes only when its refined token count is strictly lower than its own baseline. `,i.failedTaskCount,` failed this token-efficiency test; aggregate usage changed by `,c,`.`]})]}):(0,O.jsx)(`p`,{className:`training-run-placeholder`,children:`Paired task-level token evidence is unavailable for this run.`}),(0,O.jsxs)(`dl`,{className:`labs-run-detail-list`,children:[(0,O.jsx)(q,{label:`Execution`,value:o(t.status)}),(0,O.jsx)(q,{label:`Token-efficiency result`,value:s?`Passed`:`Failed`}),(0,O.jsx)(q,{label:`Efficiency passes`,value:`${i.passedTaskCount}/${i.comparedTaskCount}`}),(0,O.jsx)(q,{label:`Efficiency failures`,value:`${i.failedTaskCount}/${i.comparedTaskCount}`}),(0,O.jsx)(q,{label:`All-task foreground tokens`,value:`${i.baselineTokens.toLocaleString()} → ${i.refinedTokens.toLocaleString()}`}),(0,O.jsx)(q,{label:`Aggregate token change`,value:c}),(0,O.jsx)(q,{label:`Adaptation replay`,value:l(`adaptation`)}),(0,O.jsx)(q,{label:`Held-out replay`,value:l(`held_out`)}),(0,O.jsx)(q,{label:`Task quality`,value:`${i.baselinePassedCount}/${i.comparedTaskCount} → ${i.refinedPassedCount}/${i.comparedTaskCount}`}),(0,O.jsx)(q,{label:`Result lineage`,value:e.lineage.valid?`Valid`:`Invalid`}),(0,O.jsxs)(`div`,{children:[(0,O.jsx)(`dt`,{children:`Taskset`}),(0,O.jsx)(`dd`,{children:r?(0,O.jsx)(`button`,{className:`labs-run-taskset-link`,type:`button`,onClick:r,children:n}):n})]})]})]})}function G({title:e,baseline:t,candidate:n,format:r,summarize:i=e=>e.reduce((e,t)=>e+t,0)}){let a=t.filter(e=>e!==null),o=n.filter(e=>e!==null),s=Math.max(1,...a,...o),c=Math.max(2,t.length,n.length),l=e=>{let t=[],n=[];return e.forEach((e,r)=>{if(e===null){n.length&&t.push(n.join(` `)),n=[];return}let i=6+r/(c-1)*88,a=44-e/s*36;n.push(`${i},${a}`)}),n.length&&t.push(n.join(` `)),t},u=e=>e.flatMap((e,t)=>e===null?[]:[{x:6+t/(c-1)*88,y:44-e/s*36}]);return(0,O.jsxs)(`section`,{className:`labs-benchmark-line-chart`,children:[(0,O.jsxs)(`header`,{children:[(0,O.jsx)(`h4`,{children:e}),(0,O.jsxs)(`div`,{children:[(0,O.jsxs)(`span`,{children:[`Baseline `,r(i(a))]}),(0,O.jsxs)(`span`,{children:[`Refined `,r(i(o))]})]})]}),(0,O.jsxs)(`svg`,{"aria-label":e,role:`img`,viewBox:`0 0 100 50`,preserveAspectRatio:`none`,children:[(0,O.jsx)(`line`,{x1:`6`,x2:`94`,y1:`44`,y2:`44`}),l(t).map((e,t)=>(0,O.jsx)(`polyline`,{className:`baseline`,points:e},`baseline-${t}`)),l(n).map((e,t)=>(0,O.jsx)(`polyline`,{className:`candidate`,points:e},`candidate-${t}`)),u(t).map((e,t)=>(0,O.jsx)(`circle`,{className:`baseline`,cx:e.x,cy:e.y,r:`1`},`baseline-point-${t}`)),u(n).map((e,t)=>(0,O.jsx)(`circle`,{className:`candidate`,cx:e.x,cy:e.y,r:`1`},`candidate-point-${t}`))]})]})}function K({receipt:e,onOpenConversation:t}){let n=e.schemaVersion===`openpond.modelEvaluationReceipt.v1`?b(e.attempts??[]):e.attempts??[];return(0,O.jsxs)(`section`,{className:`labs-run-summary-card`,children:[(0,O.jsx)(`header`,{children:(0,O.jsx)(`h3`,{children:e.schemaVersion===`openpond.modelEvaluationReceipt.v1`?`Selected task results`:`Completed attempts`})}),(0,O.jsx)(`div`,{className:`training-table-wrap`,children:(0,O.jsxs)(`table`,{className:`training-data-table labs-benchmark-attempt-table`,children:[(0,O.jsx)(`thead`,{children:(0,O.jsxs)(`tr`,{children:[(0,O.jsx)(`th`,{children:`Task`}),(0,O.jsx)(`th`,{children:`Phase`}),(0,O.jsx)(`th`,{children:`Result`}),(0,O.jsx)(`th`,{children:`Score`}),(0,O.jsx)(`th`,{children:`Tokens`}),(0,O.jsx)(`th`,{children:`Work chat`})]})}),(0,O.jsx)(`tbody`,{children:n.map(e=>(0,O.jsxs)(`tr`,{children:[(0,O.jsx)(`td`,{children:we(e.taskId)}),(0,O.jsx)(`td`,{children:e.phase}),(0,O.jsx)(`td`,{children:(0,O.jsx)(_,{label:e.passed?`Passed`:`Failed`,value:e.passed?`succeeded`:`failed`})}),(0,O.jsx)(`td`,{children:e.score===null?`—`:e.score.toFixed(2)}),(0,O.jsx)(`td`,{children:e.totalTokens.toLocaleString()}),(0,O.jsx)(`td`,{children:e.sessionId?(0,O.jsx)(`button`,{className:`labs-run-taskset-link`,type:`button`,onClick:()=>t(e.sessionId),children:`Open chat`}):`Not retained`})]},`${e.phase}:${e.attemptId}`))})]})})]})}function we(e){return e.replace(/^(?:adaptation|frozen)-/,``).replaceAll(`-`,` `).replace(/^./,e=>e.toUpperCase())}function q({label:e,value:t}){return(0,O.jsxs)(`div`,{children:[(0,O.jsx)(`dt`,{children:e}),(0,O.jsx)(`dd`,{children:t})]})}function Te(e){let t=[`B`,`KB`,`MB`,`GB`,`TB`],n=e,r=0;for(;n>=1024&&r<t.length-1;)n/=1024,r+=1;return`${n.toFixed(r===0?0:2)} ${t[r]}`}function Ee({error:e,events:t,loading:n,pending:r}){let[i,o]=(0,D.useState)(!1),[s,c]=(0,D.useState)({}),[u,p]=(0,D.useState)(!0),h=(0,D.useRef)(null),g=(0,D.useRef)(null),_=(0,D.useRef)(null),v=(0,D.useMemo)(()=>De(t),[t]),y=(0,D.useMemo)(()=>t.filter(e=>s[J(e)]??Y(J(e))),[t,s]),b=v.filter(e=>s[e.id]??Y(e.id)).length,x=v.length>0&&b===v.length,S=b>0&&!x;return(0,D.useEffect)(()=>{_.current&&(_.current.indeterminate=S)},[S]),(0,D.useEffect)(()=>{if(!i)return;let e=e=>{g.current?.contains(e.target)||o(!1)},t=e=>{e.key===`Escape`&&o(!1)};return document.addEventListener(`pointerdown`,e),document.addEventListener(`keydown`,t),()=>{document.removeEventListener(`pointerdown`,e),document.removeEventListener(`keydown`,t)}},[i]),(0,D.useLayoutEffect)(()=>{let e=h.current;e&&u&&(e.scrollTop=e.scrollHeight)},[e,u,r,y]),n&&!t.length?(0,O.jsx)(`div`,{className:`training-run-placeholder`,children:`Loading run events…`}):t.length?(0,O.jsxs)(`div`,{className:`training-event-log-shell`,children:[(0,O.jsxs)(`div`,{className:`training-event-log-toolbar`,children:[(0,O.jsx)(`div`,{children:(0,O.jsx)(`strong`,{children:`Activity log`})}),(0,O.jsxs)(`div`,{className:`training-event-filter`,ref:g,children:[(0,O.jsxs)(`button`,{"aria-expanded":i,"aria-haspopup":`menu`,className:`training-event-filter-trigger`,type:`button`,onClick:()=>o(e=>!e),children:[(0,O.jsx)(f,{"aria-hidden":`true`,size:14}),`Event types`,(0,O.jsxs)(`span`,{children:[b,`/`,v.length]}),(0,O.jsx)(l,{"aria-hidden":`true`,size:13})]}),i?(0,O.jsxs)(`div`,{"aria-label":`Filter training activity by event type`,className:`training-event-filter-menu`,role:`menu`,children:[(0,O.jsxs)(`label`,{className:`training-event-filter-option select-all`,children:[(0,O.jsx)(`input`,{checked:x,ref:_,type:`checkbox`,onChange:e=>{let t=e.currentTarget.checked;c(Object.fromEntries(v.map(e=>[e.id,t])))}}),(0,O.jsx)(`span`,{className:`training-event-filter-check`,children:(0,O.jsx)(m,{"aria-hidden":`true`,size:12})}),`Select all`]}),(0,O.jsx)(`div`,{className:`training-event-filter-options`,children:v.map(e=>{let t=s[e.id]??Y(e.id);return(0,O.jsxs)(`label`,{className:`training-event-filter-option`,children:[(0,O.jsx)(`input`,{checked:t,type:`checkbox`,onChange:t=>c(n=>({...n,[e.id]:t.currentTarget.checked}))}),(0,O.jsx)(`span`,{className:`training-event-filter-check`,children:(0,O.jsx)(m,{"aria-hidden":`true`,size:12})}),(0,O.jsx)(`span`,{children:e.label}),(0,O.jsx)(`small`,{children:e.count.toLocaleString()})]},e.id)})})]}):null]})]}),(0,O.jsxs)(`div`,{"aria-label":`Training activity log`,"aria-live":`polite`,className:`training-event-log`,ref:h,role:`log`,onScroll:e=>{let t=e.currentTarget,n=t.scrollHeight-t.scrollTop-t.clientHeight;p(n<=6)},children:[(0,O.jsxs)(`div`,{"aria-hidden":`true`,className:`training-event-log-header`,children:[(0,O.jsx)(`span`,{children:`Time`}),(0,O.jsx)(`span`,{children:`Event type`}),(0,O.jsx)(`span`,{children:`Details`})]}),y.map(e=>(0,O.jsxs)(`div`,{className:`training-event-log-line ${e.type}`,children:[(0,O.jsx)(`time`,{dateTime:e.timestamp,children:a(e.timestamp)}),(0,O.jsx)(`span`,{children:Oe(J(e))}),(0,O.jsx)(`code`,{children:X(e)})]},e.id)),r&&!e?(0,O.jsxs)(`div`,{className:`training-event-log-loading`,role:`status`,children:[(0,O.jsx)(d,{"aria-hidden":`true`,className:`spin`,size:14}),(0,O.jsx)(`span`,{className:`sr-only`,children:`Waiting for new activity`})]}):null,y.length?null:(0,O.jsx)(`div`,{className:`training-event-log-empty`,children:`No events match the selected event types.`}),e?(0,O.jsxs)(`div`,{className:`training-event-log-line failure`,children:[(0,O.jsx)(`time`,{children:`—`}),(0,O.jsx)(`span`,{children:`Failure`}),(0,O.jsx)(`code`,{children:e})]}):null]})]}):(0,O.jsx)(`div`,{className:`training-run-placeholder`,children:e??`No normalized run events were recorded.`})}function J(e){return typeof e.payload.remoteEventType==`string`&&e.payload.remoteEventType.trim()?e.payload.remoteEventType:typeof e.payload.telemetryType==`string`&&e.payload.telemetryType.trim()?e.payload.telemetryType:e.type}function De(e){let t=new Map;for(let n of e){let e=J(n);t.set(e,(t.get(e)??0)+1)}return[...t.entries()].map(([e,t])=>({id:e,count:t,label:Oe(e)})).sort((e,t)=>e.label.localeCompare(t.label))}function Y(e){return e!==`metric`&&!e.endsWith(`_metric`)}function Oe(e){return e.replaceAll(`_`,` `).replace(/^./,e=>e.toUpperCase())}function X(e){let t=e.payload,n=Q(t.step),r=Q(t.maxSteps);if(typeof t.telemetryType==`string`){let e=typeof t.message==`string`?t.message:null,r=typeof t.errorCode==`string`?t.errorCode:null,i=typeof t.telemetrySource==`string`?` · ${t.telemetrySource}`:``;return`${e??t.telemetryType.replaceAll(`_`,` `)}${n==null?``:` · step ${n}`}${i}${r?` · ${r}`:``}`}if(typeof t.remoteEventType==`string`){let e=ke(t.remoteEventType),r=typeof t.remotePhase==`string`?Ae(t.remotePhase):null,i=typeof t.errorCode==`string`&&t.errorCode.trim()?t.errorCode.trim():null;return[e,r,n==null?null:`step ${n}`,i?`error ${i}`:null].filter(e=>!!e).join(` · `)}if(e.type===`start`)return typeof t.device==`string`?`Worker started on ${t.device}.`:`Worker started.`;if(e.type===`progress`&&n!=null)return r==null?`Step ${n}.`:`Step ${n} of ${r}.`;if(e.type===`metric`){let e=typeof t.metricKind==`string`?t.metricKind.replaceAll(`_`,` `):`metric`,r=[Z(`loss`,t.loss),Z(`reward`,t.meanReward??t.reward),Z(`policy loss`,t.policyLoss),Z(`value loss`,t.valueLoss),Z(`preference accuracy`,t.preferenceAccuracy,Me)].filter(e=>!!e),i=typeof t.failureClass==`string`?t.failureClass:null,a=typeof t.failureCode==`string`?t.failureCode:null,o=typeof t.metricId==`string`&&typeof t.value==`number`?` · ${t.metricId}: ${t.value.toPrecision(4)}`:``;return`${e}${n==null?``:` · step ${n}`}${o}${i?` · ${i}`:``}${a?` · ${a}`:``}${r.length?` · ${r.join(` · `)}`:``}`}if(e.type===`checkpoint`){let e=Q(t.policyVersion),n=Q(t.sizeBytes),r=[e==null?null:`policy ${e}`,n==null?null:Te(n),t.final===!0?`final`:null].filter(e=>!!e);return r.length?`Checkpoint committed · ${r.join(` · `)}.`:`Checkpoint committed.`}if(e.type===`complete`){let e=Q(t.artifactCount);return e==null?`Worker completed.`:`Worker completed with ${e} artifacts.`}return e.type===`failure`&&typeof t.message==`string`?t.message:Object.keys(t).length?JSON.stringify(t):`Recorded.`}function ke(e){return{drain:`Rollout drain`,infer:`Model inference`,materialize_checkpoint:`Checkpoint materialization`,optimizer_metric:`Optimizer metric`,provision_gpu:`GPU provisioning`,rollout_metric:`Rollout trajectory`,score_reward_model:`Reward scoring`,start_inference:`Policy server`,train_step:`Optimizer update`,upload_checkpoint:`Checkpoint upload`}[e]??je(e)}function Ae(e){return{committed:`committed`,completed:`completed`,eligible:`recorded`,failed:`failed`,running:`running`,succeeded:`succeeded`}[e]??je(e).toLocaleLowerCase()}function je(e){return e.replaceAll(`_`,` `).replace(/^./,e=>e.toUpperCase())}function Z(e,t,n=e=>new Intl.NumberFormat(void 0,{maximumFractionDigits:4}).format(e)){let r=Q(t);return r==null?null:`${e} ${n(r)}`}function Q(e){return typeof e==`number`&&Number.isFinite(e)?e:null}function Me(e){return`${(e*100).toFixed(1)}%`}function Ne({connection:e,detailTab:t,onDetailTabChange:n,selectedEntryKey:a,workproduct:l,runs:u,training:d,onOpenDataset:f,onOpenConversation:p,detailKind:m}){let h=d.payload,[_,v]=(0,D.useState)(null),b=(0,D.useMemo)(()=>S(l,u,h),[u,h,l]),T=(0,D.useMemo)(()=>ne(l,u,h),[u,h,l]),E=(0,D.useMemo)(()=>ae(l,h),[h,l]),A=(0,D.useMemo)(()=>ee(l,u,h),[u,h,l]),j=(0,D.useMemo)(()=>new Map(A.map(e=>[e.id,e])),[A]),M=(0,D.useMemo)(()=>te(b,T),[b,T]).find(e=>e.key===a)??null,ce=a.startsWith(`model-run:`)?a.slice(10):null,N=ce?E.find(e=>e.id===ce)??null:M?.job?E.find(e=>e.id===M.job?.metadata.modelRunId)??null:null,P=M?.job??(N?h?.jobs.find(e=>e.id===N.id||e.metadata.modelRunId===N.id)??null:null),ue=h?.comparisonSeriesEntries.find(e=>e.modelRunId===N?.id||e.modelRunId===P?.metadata.modelRunId)??null,de=typeof P?.metadata.parentHostedArtifactId==`string`?`Prior hosted run`:Ue(h,ue),fe=(0,D.useMemo)(()=>{let e=ie(b,T,E);return x(e,N?.kind===`evaluation`?`evals`:`training`)},[b,E,N?.kind,T]),pe=fe.findIndex(e=>e.key===a||P!==null&&e.job?.id===P.id||N!==null&&e.lifecycleRun?.id===N.id),F=pe<0?null:fe.length-pe,I=M?.version??(N?.adapterArtifactLineageId?T.find(e=>e.lineage.id===N.adapterArtifactLineageId)??null:null),ge=(N?h?.modelVersions.find(e=>e.id===N.modelVersionId)?.baseModel.modelId:null)??h?.modelVersions.filter(e=>e.modelId===l.id&&e.kind===`base_reference`).sort((e,t)=>t.version-e.version)[0]?.baseModel.modelId??(typeof P?.metadata.baseModelId==`string`?P.metadata.baseModelId:null),_e=I?.lineage.frozenEvaluationArtifactId??null,L=I?.plan??(P?j.get(P.planId)??null:null),R=I?.taskset??re(h).find(e=>e.id===(N?.taskset.id??L?.tasksetId))??null,z=R?.name??(typeof P?.metadata.tasksetId==`string`?P.metadata.tasksetId:`Unavailable`),B=N?.method??L?.recipe.method??Re(P?.metadata),V=ve(e,P?.id??null,P?.status??null,t===`evaluation`),H=P?h?.rolloutReceipts.filter(e=>e.jobId===P.id)??[]:[],U=V.detail?.managedEvidence??null,be=V.detail?.events.some(e=>e.type===`metric`&&e.payload.metricKind===`rollout_trajectory`)??!1,W=oe({lifecycleRun:N,job:V.detail?.job??P}),xe=y(W),Se=N?.status===`failed`&&P?.metadata.phase===`artifact_collection_failed`,Ce=B===`grpo`,G=Fe(P?.metadata),K=(L&&(L.recipe.method===`sft`||L.recipe.method===`dpo`||L.recipe.method===`grpo`)?L.recipe.optimizer.maxSteps:null)??U?.progress.targetOptimizerSteps??null,we=K!==null&&typeof P?.metadata.progress==`number`&&Number.isFinite(P.metadata.progress)?Math.floor(Math.max(0,P.metadata.progress)*K):0,q=Math.max(0,G?.optimizerUpdatesApplied??0,we,V.detail?.policyMetrics.reduce((e,t)=>Math.max(e,t.step),0)??0,V.detail?.stepMetrics.reduce((e,t)=>Math.max(e,t.step),0)??0,U?.progress.committedOptimizerSteps??0),J=Pe(Ce&&G?G.groupsCompleted:q,Ce&&G?G.groupsTarget:K),De=m===`run`?{id:`details`,label:`Details`}:{id:`overview`,label:`Overview`},Y=[...m===`run`&&P?[{id:`metrics`,label:`Metrics`}]:[],De,...m===`version`&&P?[{id:`metrics`,label:`Metrics`}]:[],{id:`evaluation`,label:`Evaluation`},...H.length||be?[{id:`rollouts`,label:`Rollouts`}]:[],{id:`artifacts`,label:`Artifacts`},...P?[{id:`activity`,label:`Activity`}]:[]],Oe=t,X=Y.some(e=>e.id===Oe)?Oe:Y[0].id;return!M&&!N?(0,O.jsx)(`div`,{className:`labs-model-version-detail`,children:(0,O.jsx)(`div`,{className:`training-run-placeholder`,children:`This training attempt is no longer available.`})}):N?.kind===`evaluation`?(0,O.jsx)(Ie,{run:N,runNumber:F,tasksetName:R?.name??`Unavailable`,onOpenTaskset:R?()=>f(R.id):void 0,onOpenConversation:p}):(0,O.jsxs)(`div`,{className:`labs-model-version-detail`,children:[(0,O.jsx)(w,{title:N||P?F?`Run ${F}`:`Run details`:I?`Version ${I.number}`:`Run details`,description:`${i(B)} on ${Le(L,ge)}`,status:(0,O.jsx)(C,{status:W}),actions:xe&&(N||P)?(0,O.jsx)(`button`,{type:`button`,className:`training-button secondary`,disabled:d.busyAction!==null||W===`cancelling`,onClick:async()=>{let e=N?.id??P.id,t=N?await d.actions.cancelModelRun(N.id):await d.actions.cancelJob(P.id);v({runId:e,message:t?`Cancellation requested. Waiting for the training service to confirm.`:`Cancellation could not be requested. Try again.`})},children:W===`cancelling`?`Cancelling…`:d.busyAction===`cancel-model-run`||d.busyAction===`cancel-job`?`Requesting cancellation…`:`Cancel run`}):Se&&N?(0,O.jsx)(`button`,{type:`button`,className:`training-button secondary`,disabled:d.busyAction!==null,onClick:async()=>{let e=await d.actions.retryModelRunCollection(N.id);v({runId:N.id,message:e?.state===`succeeded`?`Outputs collected from the completed training job.`:`Collection failed. The training job and its outputs are retained.`})},children:d.busyAction===`retry-model-run-collection`?`Collecting…`:`Retry collection`}):void 0,metrics:[{label:Ce?`Training tasks`:`Training steps`,value:J.value},{label:`Final reward`,value:Be(N?.reward?.raw??U?.reward.finalMean??null)},{label:`Duration`,value:N?s(N.startedAt,r(N.status,N.completedAt,N.updatedAt)):P?s(P.startedAt,r(P.status,P.completedAt,P.updatedAt)):`Not recorded`},{label:`Output`,value:I?`Version ${I.number}`:`No version`},{label:`Based on`,value:de},{label:`Taskset`,value:z,onSelect:R?()=>f(R.id):void 0,ariaLabel:R?`Open Taskset ${R.name}`:void 0}]}),_&&_.runId===(N?.id??P?.id)?(0,O.jsx)(`p`,{role:`status`,children:_.message}):null,(0,O.jsxs)(`section`,{className:`labs-run-detail-tabs`,children:[(0,O.jsx)(`div`,{className:`training-detail-tabs`,role:`tablist`,"aria-label":`Run details`,children:Y.map(e=>(0,O.jsx)(`button`,{"aria-controls":`run-detail-panel-${e.id}`,"aria-selected":X===e.id,className:X===e.id?`active`:void 0,id:`run-detail-tab-${e.id}`,role:`tab`,type:`button`,onClick:()=>n(e.id),children:e.label},e.id))}),(0,O.jsxs)(`div`,{"aria-labelledby":`run-detail-tab-${X}`,className:`labs-run-detail-tab-panel`,id:`run-detail-panel-${X}`,role:`tabpanel`,children:[X===De.id?(0,O.jsx)(ye,{baseModel:Le(L,ge),compute:N?c(N.destinationId??``):L?c(L.destinationId):P?c(P.destinationId):`Not recorded`,duration:N?s(N.startedAt,r(N.status,N.completedAt,N.updatedAt)):P?s(P.startedAt,r(P.status,P.completedAt,P.updatedAt)):`Not recorded`,failure:P?.error??N?.failure??null,configuration:[...He(h,ue),...Ve(L)],evidence:U,method:i(B),output:I?`Version ${I.number}`:`No version created`,reward:N?.reward?.raw??U?.reward.finalMean??null,status:N?o(N.status):P?o(P.status):`Imported`,statusValue:N?.status??P?.status??`imported`,taskset:z,telemetry:N?.receipt?.schemaVersion===`openpond.modelRunReceipt.v1`?N.receipt.telemetry:null,title:N||P?F?`Run ${F}`:`Run details`:I?`Version ${I.number}`:`Run details`,versionStatus:I?I.current?`Active`:`Available`:null,onOpenTaskset:R?()=>f(R.id):void 0,showHeader:!1}):null,X===`metrics`&&P?(0,O.jsx)(g,{detail:V.detail,error:V.error,loading:V.loading,taskset:R}):null,X===`artifacts`?(0,O.jsxs)(`dl`,{className:`training-configuration-list`,children:[(0,O.jsx)($,{label:`Training attempt`,value:P?.id??`Provider import`}),(0,O.jsx)($,{label:`Taskset`,value:z}),(0,O.jsx)($,{label:`Prepared data`,value:P?.bundleHash??`Provider managed`}),(0,O.jsx)($,{label:`Version ID`,value:I?.lineage.id??`No version created`}),U?.checkpoint?(0,O.jsxs)(O.Fragment,{children:[(0,O.jsx)($,{label:`Checkpoint`,value:U.checkpoint.id}),(0,O.jsx)($,{label:`Checkpoint size`,value:U.checkpoint.sizeBytes==null?`Not reported`:Te(U.checkpoint.sizeBytes)}),(0,O.jsx)($,{label:`Checkpoint hash`,value:U.checkpoint.sha256??`Not reported`})]}):null,U?.canonicalPublication.state?(0,O.jsx)($,{label:`Registry publication`,value:Ge(U.canonicalPublication.state)}):null]}):null,X===`evaluation`?(0,O.jsxs)(`div`,{className:`training-run-evaluation`,children:[!U?.evaluations.length||V.detail?.evaluation?(0,O.jsx)(le,{detail:V.detail,loading:V.loading,pending:xe}):null,!V.detail?.evaluation&&U?.evaluations.length?(0,O.jsx)(`dl`,{className:`labs-inline-facts`,children:U.evaluations.map(e=>(0,O.jsxs)(D.Fragment,{children:[(0,O.jsx)($,{label:e.kind===`baseline`?`Baseline score`:`Candidate score`,value:e.score==null?`Not reported`:e.score.toFixed(6)}),(0,O.jsx)($,{label:e.kind===`baseline`?`Baseline coverage`:`Candidate coverage`,value:e.taskCount==null?`Not reported`:e.targetTaskCount==null?`${e.taskCount} ${e.taskCount===1?`task`:`tasks`} scored`:`${e.taskCount} of ${e.targetTaskCount} ${e.targetTaskCount===1?`task`:`tasks`} scored`})]},`${e.kind}:${e.policyVersion}`))}):null,I&&P?.destinationId===`openpond_managed`?(0,O.jsx)(me,{connection:e,lineageId:I.lineage.id,onSaved:d.refresh},I.lineage.id):null,!I&&P?.metadata.source===`hosted_model_project_import`&&P.status===`succeeded`?(0,O.jsx)(me,{connection:e,jobId:P.id,onSaved:d.refresh},`candidate:${P.id}`):null,P&&U?.evaluations.length?(0,O.jsx)(he,{connection:e,jobId:P.id,evaluations:U.evaluations},P.id):null,_e?(0,O.jsx)(`button`,{className:`training-button secondary`,type:`button`,onClick:()=>void d.actions.downloadArtifact(_e),children:`Download evaluation receipt`}):null]}):null,X===`rollouts`?H.length?(0,O.jsx)(se,{receipts:H}):(0,O.jsx)(k,{events:V.detail?.events??[]}):null,X===`activity`&&P?(0,O.jsx)(Ee,{error:P.error??V.error,events:V.detail?.events??[],loading:V.loading,pending:xe}):null]})]})]})}function Pe(e,t){let n=Math.max(0,Math.floor(e));if(t===null||!Number.isFinite(t)||t<=0)return{value:n.toLocaleString(),hint:n===1?`completed update`:`completed updates`};let r=Math.max(1,Math.floor(t));return{value:`${n.toLocaleString()} / ${r.toLocaleString()}`,hint:`completed / planned`}}function Fe(e){let t=e?.rolloutProgress;if(!t||typeof t!=`object`||Array.isArray(t))return null;let n=t;return[`groupsCompleted`,`groupsTarget`,`optimizerUpdatesApplied`,`optimizerUpdatesSkipped`].some(e=>typeof n[e]!=`number`||!Number.isInteger(n[e])||n[e]<0)?null:{groupsCompleted:n.groupsCompleted,groupsTarget:n.groupsTarget,optimizerUpdatesApplied:n.optimizerUpdatesApplied,optimizerUpdatesSkipped:n.optimizerUpdatesSkipped}}function Ie({run:e,runNumber:t,tasksetName:n,onOpenTaskset:r,onOpenConversation:i}){let[s,c]=(0,D.useState)(`results`),l=e.receipt?.schemaVersion===`openpond.modelEvaluationReceipt.v1`?e.receipt:null,u=e.receipt?.schemaVersion===`openpond.modelEvaluationStopReceipt.v1`?e.receipt:null,d=l?E(l):null,f=l?T(l):null;return(0,O.jsxs)(`div`,{className:`labs-model-version-detail labs-model-evaluation-detail`,children:[(0,O.jsx)(w,{title:t?`Eval ${t}`:`Evaluation run`,description:`Harness Refiner benchmark · ${n}`,status:(0,O.jsx)(_,{label:l?f?.comparedTaskCount?`${f.passedTaskCount}/${f.comparedTaskCount} passed`:`No paired result`:u?`Inconclusive`:o(e.status),value:l?f?.comparedTaskCount&&f.passed?`succeeded`:`neutral`:u?.terminalClassification??e.status}),metrics:[{label:`Baseline pass rate`,value:l?`${(l.quality.baselinePassRate*100).toFixed(1)}%`:`—`},{label:`Candidate pass rate`,value:l?`${(l.quality.candidatePassRate*100).toFixed(1)}%`:`—`},{label:`Scenarios`,value:l?.attempts?.length??0},{label:`Execution`,value:o(e.status)}]}),(0,O.jsx)(`div`,{className:`training-detail-tabs`,role:`tablist`,"aria-label":`Evaluation details`,children:[`results`,`scenarios`,`activity`].map(e=>(0,O.jsx)(`button`,{className:s===e?`active`:void 0,"aria-selected":s===e,role:`tab`,type:`button`,onClick:()=>c(e),children:e===`results`?`Results`:e===`scenarios`?`Scenarios`:`Activity`},e))}),s===`results`&&l?(0,O.jsxs)(O.Fragment,{children:[(0,O.jsx)(v,{series:[{id:`baseline`,label:`Held-out baseline`,inputTokens:d.baseline.inputTokens,outputTokens:d.baseline.outputTokens,tokens:d.baseline.totalTokens,passRate:l.quality.baselinePassRate,costUsd:d.baseline.costUsd},{id:`candidate`,label:`Held-out refined`,inputTokens:d.candidate.inputTokens,outputTokens:d.candidate.outputTokens,tokens:d.candidate.totalTokens,passRate:l.quality.candidatePassRate,costUsd:d.candidate.costUsd}]}),(0,O.jsx)(v,{series:[{id:`adaptation-baseline`,label:`Adaptation baseline`,inputTokens:d.adaptation.inputTokens,outputTokens:d.adaptation.outputTokens,tokens:d.adaptation.totalTokens,passRate:l.quality.adaptationBaselinePassRate,costUsd:d.adaptation.costUsd},{id:`adaptation-candidate`,label:`Adaptation replay`,inputTokens:d.candidate_adaptation.inputTokens,outputTokens:d.candidate_adaptation.outputTokens,tokens:d.candidate_adaptation.totalTokens,passRate:l.quality.adaptationCandidatePassRate,costUsd:d.candidate_adaptation.costUsd}]}),(0,O.jsx)(Ce,{receipt:l,run:e,tasksetName:n,onOpenTaskset:r})]}):s===`results`&&u?(0,O.jsx)(W,{receipt:u,onOpenConversation:i}):s===`results`?(0,O.jsx)(`section`,{className:`labs-run-summary-card`,children:e.failure?(0,O.jsx)(`p`,{className:`training-run-placeholder`,children:e.failure}):(0,O.jsx)(xe,{run:e})}):null,s===`scenarios`&&l?(0,O.jsxs)(O.Fragment,{children:[(l.attempts??[]).length?(0,O.jsx)(Se,{receipt:l}):null,(l.attempts??[]).length?(0,O.jsx)(K,{receipt:l,onOpenConversation:i}):(0,O.jsx)(`div`,{className:`training-run-placeholder`,children:`No per-scenario attempts were retained.`})]}):null,s===`activity`?(0,O.jsxs)(`dl`,{className:`labs-inline-facts`,children:[(0,O.jsx)($,{label:`Run ID`,value:e.id}),(0,O.jsx)($,{label:`Status`,value:o(e.status)}),(0,O.jsx)($,{label:`Started`,value:e.startedAt?a(e.startedAt):`Not started`}),(0,O.jsx)($,{label:`Completed`,value:e.completedAt?a(e.completedAt):`In progress`})]}):null]})}function Le(e,t=null){if(!e)return t?ze(t):`Not recorded`;let n=e.recipe.method===`sft`||e.recipe.method===`grpo`?e.recipe.baseModel:e.recipe.method===`dpo`?e.recipe.policyModel:e.recipe.method===`ppo`?e.recipe.policyOptimization.policyModel:null;return n?ze(n.id):`Not recorded`}function Re(e){let t=e?.trainingMethod;return t===`sft`||t===`dpo`||t===`grpo`||t===`ppo`||t===`sdft`||t===`opd`||t===`opsd`||t===`sdpo`?t:null}function ze(e){return e.split(`/`).at(-1)??e}function Be(e){return e===null?`Not reported`:e.toFixed(4)}function Ve(e){if(!e)return[];let t=e.recipe;return t.method===`grpo`?[{label:`LoRA rank`,value:String(t.lora.rank)},{label:`Learning rate`,value:We(t.optimizer.learningRate)},{label:`Optimizer steps`,value:String(t.optimizer.maxSteps)},{label:`Rollouts per update`,value:String(t.rollout.groupSize)},{label:`Rollout concurrency`,value:String(t.rollout.concurrency)},{label:`Max output tokens`,value:t.rollout.maxOutputTokens.toLocaleString()},{label:`Temperature`,value:String(t.rollout.temperature)},{label:`Top P`,value:String(t.rollout.topP)}]:t.method===`sft`||t.method===`dpo`?[{label:`LoRA rank`,value:String(t.lora.rank)},{label:`Learning rate`,value:We(t.optimizer.learningRate)},{label:`Optimizer steps`,value:String(t.optimizer.maxSteps)}]:[]}function He(e,t){if(!e||!t)return[];if(t.parent.kind===`base_model`)return[{label:`Starting checkpoint`,value:`Frozen base · ${t.parent.revision}`},{label:`Optimizer`,value:`Fresh`}];let n=e.modelVersions.find(e=>e.id===t.parent.id);return[{label:`Starting checkpoint`,value:[e.comparisonSeriesEntries.find(e=>e.seriesId===t.seriesId&&e.modelVersionId===t.parent.id)?.label??null,n?`Version ${n.version}`:t.parent.id].filter(e=>!!e).join(` · `)},{label:`Optimizer`,value:`Fresh`}]}function Ue(e,t){if(!e||!t||t.parent.kind===`base_model`)return`Frozen base`;let n=e.comparisonSeriesEntries.find(e=>e.seriesId===t.seriesId&&e.modelVersionId===t.parent.id),r=e.modelVersions.find(e=>e.id===t.parent.id);return[n?.label??`Prior checkpoint`,r?`Version ${r.version}`:null].filter(e=>!!e).join(` · `)}function We(e){return e===0?`0`:e.toExponential(2)}function $({label:e,value:t}){return(0,O.jsxs)(`div`,{children:[(0,O.jsx)(`dt`,{children:e}),(0,O.jsx)(`dd`,{children:t})]})}function Ge(e){return e?e.replaceAll(`_`,` `).replace(/^./,e=>e.toUpperCase()):`Pending`}export{Ne as LabModelVersionDetailPage};

package/dist/web/assets/LabSkillSidebar-gjbl9Nlv.js

@@ -1 +0,0 @@
-import{i as e}from"./rolldown-runtime-Dd_uD5pT.js";import{g as t,m as n}from"./index-BsPUniZy.js";import"./icons-kaq6k88m.js";import{L as r,R as i}from"./training-model-data-CmVo80ZR.js";import{t as a}from"./image-C7HVlg3X.js";import{t as o}from"./maximize-2-DUWvLap6.js";import{Bn as s,Mn as c,_n as l,rr as u,vn as d}from"./App-Yy3l2sdD.js";var f=e(t(),1),p=n();function m({connection:e,expanded:t,selection:n,onClose:i,onResizeStart:s,onToggleExpanded:m}){let v=`${n.scope}:${n.name}`,[y,b]=(0,f.useState)({selectionKey:v,path:n.files[0]??`SKILL.md`}),x=y.selectionKey===v&&n.files.includes(y.path)?y.path:n.files[0]??`SKILL.md`,[S,C]=(0,f.useState)(null),[w,T]=(0,f.useState)(null),[E,D]=(0,f.useState)(!1);(0,f.useEffect)(()=>{if(!e){C(null),T(`Connect to the local OpenPond server to load this skill source.`);return}let t=!1;return D(!0),T(null),C(null),u.skillSourceFile(e,n.scope,n.name,x).then(e=>{t||C(e)}).catch(e=>{t||T(e instanceof Error?e.message:String(e))}).finally(()=>{t||D(!1)}),()=>{t=!0}},[e,x,n.name,n.scope]);let O=x.split(`/`).at(-1)??x;return(0,p.jsxs)(`aside`,{className:`workspace-diff-panel native-skill-sidebar lab-skill-sidebar ${t?`expanded`:``}`,"aria-label":`Skill package source: ${n.name}`,children:[t?null:(0,p.jsx)(`div`,{className:`workspace-diff-resize-handle`,role:`separator`,"aria-orientation":`vertical`,"aria-label":`Resize skill source`,onPointerDown:s}),(0,p.jsxs)(`div`,{className:`workspace-diff-topbar native-skill-sidebar-topbar`,children:[(0,p.jsx)(`div`,{className:`workspace-diff-tabs`,role:`tablist`,"aria-label":`Selected skill file`,children:(0,p.jsxs)(`button`,{className:`workspace-diff-tab active`,type:`button`,role:`tab`,"aria-selected":!0,children:[(0,p.jsx)(g,{path:x}),(0,p.jsx)(`span`,{children:O})]})}),(0,p.jsxs)(`div`,{className:`workspace-diff-toolbar-actions`,children:[(0,p.jsx)(`button`,{className:`diff-icon-button`,type:`button`,title:t?`Dock sidebar`:`Expand sidebar`,"aria-label":t?`Dock sidebar`:`Expand sidebar`,onClick:m,children:t?(0,p.jsx)(r,{size:14}):(0,p.jsx)(o,{size:14})}),(0,p.jsx)(`button`,{className:`diff-icon-button`,type:`button`,title:`Close skill source`,"aria-label":`Close skill source`,onClick:i,children:(0,p.jsx)(d,{size:14})})]})]}),(0,p.jsxs)(`div`,{className:`lab-skill-source-layout`,children:[(0,p.jsxs)(`nav`,{className:`lab-skill-file-list`,"aria-label":`${n.name} package files`,children:[(0,p.jsx)(`div`,{className:`lab-skill-file-list-heading`,children:`Package files`}),n.files.map(e=>(0,p.jsxs)(`button`,{"aria-current":e===x?`page`:void 0,className:e===x?`active`:void 0,type:`button`,title:e,onClick:()=>b({selectionKey:v,path:e}),children:[(0,p.jsx)(g,{path:e}),(0,p.jsx)(`span`,{children:e})]},e))]}),(0,p.jsxs)(`div`,{className:`native-skill-sidebar-body lab-skill-source-body`,children:[(0,p.jsxs)(`header`,{className:`native-skill-source-header`,children:[(0,p.jsxs)(`div`,{className:`lab-skill-source-title-row`,children:[(0,p.jsx)(`h2`,{children:n.name}),(0,p.jsx)(`span`,{children:h(n.scope)})]}),(0,p.jsx)(`p`,{children:n.description}),(0,p.jsx)(`code`,{children:n.packagePath})]}),E?(0,p.jsxs)(`div`,{className:`lab-skill-source-state`,children:[(0,p.jsx)(c,{className:`spin`,size:16}),` Loading source…`]}):w?(0,p.jsx)(l,{error:w}):S?.isBinary?(0,p.jsxs)(`div`,{className:`lab-skill-source-state`,children:[(0,p.jsx)(a,{size:18}),`Binary asset · `,_(S.byteSize)]}):(0,p.jsx)(`pre`,{className:`native-skill-markdown-source`,"aria-label":`${O} source`,children:(0,p.jsx)(`code`,{children:S?.content??``})})]})]})]})}function h(e){return e===`codex`?`Codex skill`:e===`extension`?`GitHub extension`:`OpenPond skill`}function g({path:e}){let t=e.split(`.`).at(-1)?.toLowerCase();return[`avif`,`gif`,`jpeg`,`jpg`,`png`,`webp`].includes(t??``)?(0,p.jsx)(a,{size:14}):[`md`,`txt`,`yaml`,`yml`].includes(t??``)?(0,p.jsx)(s,{size:14}):(0,p.jsx)(i,{size:14})}function _(e){return e<1024?`${e} B`:e<1048576?`${(e/1024).toFixed(1)} KB`:`${(e/1048576).toFixed(1)} MB`}export{m as LabSkillSidebar};