evals-lab 0.1.4 → 0.3.0

package/lab/web/dist/gallery.html

@@ -5,10 +5,10 @@
   <meta name="viewport" content="width=device-width, initial-scale=1">
   <meta name="color-scheme" content="dark light">
   <title>Components</title>
-  <script type="module" crossorigin src="/assets/gallery-DipkRvqJ.js"></script>
-  <link rel="modulepreload" crossorigin href="/assets/tokens-B9intIuT.js">
-  <link rel="stylesheet" crossorigin href="/assets/tokens-s6I-RMVq.css">
-  <link rel="stylesheet" crossorigin href="/assets/gallery-o7c4lfpn.css">
+  <script type="module" crossorigin src="/assets/gallery-DRBlZ8mP.js"></script>
+  <link rel="modulepreload" crossorigin href="/assets/tokens-iGpvbD5R.js">
+  <link rel="stylesheet" crossorigin href="/assets/tokens-BEn6_hVz.css">
+  <link rel="stylesheet" crossorigin href="/assets/gallery-DFeJkfUw.css">
 </head>
 <body>
   <!-- The fixtures gallery: every component in every state, for the eye and

package/lab/web/dist/index.html

@@ -5,10 +5,10 @@
   <meta name="viewport" content="width=device-width, initial-scale=1">
   <meta name="color-scheme" content="dark light">
   <title>Evals Lab</title>
-  <script type="module" crossorigin src="/assets/main-61NS6C4m.js"></script>
-  <link rel="modulepreload" crossorigin href="/assets/tokens-B9intIuT.js">
-  <link rel="stylesheet" crossorigin href="/assets/tokens-s6I-RMVq.css">
-  <link rel="stylesheet" crossorigin href="/assets/main-DjQQums6.css">
+  <script type="module" crossorigin src="/assets/main-DMpQmW8l.js"></script>
+  <link rel="modulepreload" crossorigin href="/assets/tokens-iGpvbD5R.js">
+  <link rel="stylesheet" crossorigin href="/assets/tokens-BEn6_hVz.css">
+  <link rel="stylesheet" crossorigin href="/assets/main-wfqC6HcM.css">
 </head>
 <body>
   <div id="root"></div>

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "evals-lab",
-  "version": "0.1.4",
+  "version": "0.3.0",
   "description": "A browser bench for grading what a model answers: two wordings side by side over a set of files, scored against answers fixed in advance.",
   "license": "SEE LICENSE IN LICENSE.md",
   "author": "Grainbox Limited",
@@ -9,7 +9,8 @@
   },
   "files": [
     "bin/",
-    "lab/"
+    "lab/",
+    "CHANGELOG.md"
   ],
   "engines": {
     "node": ">=22.18"

package/lab/web/dist/assets/gallery-DipkRvqJ.js

@@ -1,3 +0,0 @@
-import{R as e,T as t,X as n,Y as r,_ as i,c as a,d as o,f as s,g as c,h as l,m as u,n as d,r as f,v as p}from"./tokens-B9intIuT.js";var m=n(),h=r(),g=[{id:`text`,label:`Plain text`,settings:[]},{id:`list`,label:`List`,settings:[{key:`parse`,label:`Parse as`,type:`select`,choices:[{value:`csv`,label:`CSV`},{value:`lines`,label:`One a line`},{value:`json`,label:`JSON array`}]},{key:`preamble`,label:`Ignore preamble`,type:`checkbox`}]}],_=[{id:`case`,label:`Case`,accepts:[`list`],options:[{key:`mode`,label:`Mode`,type:`select`,choices:[{value:`LOWER`,label:`lower`},{value:`SENTENCE`,label:`Sentence`},{value:`TITLE`,label:`Title`},{value:`UPPER`,label:`UPPER`},{value:`NONE`,label:`As replied`}]}],defaults:()=>({type:`case`,mode:`LOWER`})},{id:`join`,label:`Join`,accepts:[`list`],options:[{key:`with`,label:`Separator`,type:`text`}],defaults:()=>({type:`join`,with:`, `})}],v=[{name:`language`,type:`value`,value:`English UK`},{name:`geo-location`,type:`value`,value:`Hamilton, Waikato, New Zealand`},{name:`geo`,type:`block`,enabled:!1}],y=[{path:`group/open`,title:`Group, open`},{path:`group/closed`,title:`Group, closed with a note`},{path:`group/no-menu`,title:`Group without actions`},{path:`group/sub`,title:`Sub-group, level 3`},{path:`group/depth`,title:`Group, three levels deep`},{path:`group/renaming`,title:`Group, renaming`},{path:`menu/disabled`,title:`Menu with a disabled item`},{path:`repeater/empty`,title:`Repeater, empty`},{path:`repeater/min`,title:`Repeater at min 1`},{path:`repeater/between`,title:`Repeater, two of four`},{path:`repeater/max`,title:`Repeater at max`},{path:`props/controls`,title:`Property table, every control`},{path:`props/required`,title:`Property table, required only`},{path:`props/long`,title:`Property table, long values`},{path:`props/widths`,title:`Property table, a width per value`},{path:`props/secret`,title:`Property table, a key held and not held`},{path:`props/code`,title:`Property table, an endpoint and a payload`},{path:`table/sortable`,title:`Table, sortable headers and a row menu`},{path:`table/paged`,title:`Table, paged ten at a time`},{path:`table/filter`,title:`Table with a filter bar`},{path:`table/filter-on`,title:`Table, filtered, with a scope`},{path:`edit/form`,title:`Edit modal, a profile`},{path:`edit/plain`,title:`Edit modal, one section and no menu`},{path:`edit/long`,title:`Edit modal, a form that scrolls`},{path:`edit/status`,title:`Edit modal, a status that is a refusal`},{path:`view/tokens`,title:`Token mappings`},{path:`view/output`,title:`Output with Case`},{path:`view/output-text`,title:`Output, plain text`},{path:`view/output-two`,title:`Output with two modifiers`},{path:`view/output-twice`,title:`Output with the same modifier twice`}],b=e(),x=()=>{},S=(...e)=>e.map(e=>({label:e,icon:e===`Remove`?p.remove:e===`Clone`?p.clone:p.rename,onSelect:x,destructive:e===`Remove`}));function C({start:e,min:t,max:n}){let[r,i]=(0,m.useState)(e);return(0,b.jsx)(l,{noun:`scenario`,items:r,...t===void 0?{}:{min:t},...n===void 0?{}:{max:n},onAdd:()=>i([...r,`Scenario ${r.length+1}`]),onRemove:e=>i(r.filter((t,n)=>n!==e)),onClone:e=>i([...r.slice(0,e+1),`${r[e]} copy`,...r.slice(e+1)]),onRename:x,empty:`No scenarios yet.`,renderItem:(e,t,n)=>(0,b.jsx)(c,{id:`scenario-${t}`,title:e,level:3,actions:n,children:(0,b.jsx)(`p`,{className:`note`,children:`A Setup profile and a prompt for each job.`})},`${e}-${t}`)})}function w(){let[e,t]=(0,m.useState)(v),n=(n,r)=>t(e.map(e=>e.name===n?{...e,...r}:e));return(0,b.jsx)(d,{mappings:e,onValue:(e,t)=>n(e,{value:t}),onEnabled:(e,t)=>n(e,{enabled:t}),onEdit:x,onRemove:n=>t(e.filter(e=>e.name!==n)),onAdd:()=>t([...e,{name:`token${e.length+1}`,type:`value`,value:``}])})}function T({start:e}){let[t,n]=(0,m.useState)(e);return(0,b.jsx)(f,{id:`output`,out:t,outputKinds:g,modifiers:_,onKind:e=>n({kind:e,modifiers:t.modifiers.filter(t=>_.find(e=>e.id===t.type)?.accepts.includes(e))}),onModifier:(e,r)=>n({...t,modifiers:t.modifiers.map((t,n)=>n===e?r:t)}),onAddModifier:e=>n({...t,modifiers:[...t.modifiers,e]}),onRemoveModifier:e=>n({...t,modifiers:t.modifiers.filter((t,n)=>n!==e)})})}function E(){let[e,t]=(0,m.useState)(`Job 1`),[n,r]=(0,m.useState)(!0);return(0,b.jsx)(c,{id:`renaming`,title:e,actions:[{label:`Rename`,icon:p.rename,onSelect:()=>r(!0)}],...n?{rename:{value:e,onCommit:n=>{t(n||e),r(!1)},onCancel:()=>r(!1)}}:{},children:(0,b.jsx)(`p`,{className:`note`,children:`Enter or leaving keeps the name; Escape cancels.`})})}function D({sections:e,menu:n,status:r,statusBad:i,outcome:a}){return(0,b.jsx)(o,{title:`Lab Ollama`,...r?{status:r}:{},...i?{statusBad:!0}:{},...n?{menu:n}:{},sections:e,actions:[{label:`Test connection`,icon:t,onSelect:x,outcome:a?{label:a,onOpen:x}:void 0}],onClose:x})}var O=()=>[{key:`name`,label:`Name`,required:!0,control:{type:`text`,value:`Lab Ollama`,onChange:x}},{key:`type`,label:`Type`,required:!0,control:{type:`select`,value:`ollama`,onChange:x,options:[{value:`ollama`,label:`Ollama (Local)`},{value:`anthropic`,label:`Anthropic`}]}},{key:`url`,label:`Address`,control:{type:`text`,value:``,placeholder:`the lab's Ollama`,onChange:x}},{key:`key`,label:`Key`,control:{type:`secret`,value:``,onChange:x}},{key:`model`,label:`Model`,control:{type:`text`,value:`minicpm-v4.6:latest`,onChange:x}}],k=()=>[{key:`seed`,label:`Seed`,control:{type:`text`,value:``,placeholder:`1234`,marker:`default`,onChange:x}},{key:`nPredict`,label:`Reply tokens`,control:{type:`text`,value:``,size:`num`,marker:`server`,onChange:x}},{key:`temperature`,label:`Temperature`,control:{type:`text`,value:`0.3`,onChange:x}},{key:`repeatPenalty`,label:`Repeat penalty`,control:{type:`text`,value:``,size:`num`,marker:`server`,onChange:x}}],A=()=>[{key:`endpoint`,label:`Endpoint`,required:!0,control:{type:`code`,text:`http://127.0.0.1:11434/api/chat`}},{key:`prompt`,label:`Prompt`,required:!0,control:{type:`static`,text:`(prompt)`}},{key:`model`,label:`model`,control:{type:`code`,text:`minicpm-v4.6:latest`}},{key:`options`,label:`options`,control:{type:`code`,text:`{"seed":1234,"temperature":0.3}`}}],j={key:`kind`,label:`Kind`,required:!0,control:{type:`select`,value:`list`,onChange:x,options:g.map(e=>({value:e.id,label:e.label}))}},M={"group/open":()=>(0,b.jsx)(c,{id:`g`,title:`Tests`,actions:S(`Change`,`Remove`),children:(0,b.jsx)(`p`,{className:`note`,children:`The graded set.`})}),"group/closed":()=>(0,b.jsx)(c,{id:`g`,title:`Content`,open:!1,note:`Sample library, 63 files`,actions:S(`Change`,`Remove`)}),"group/no-menu":()=>(0,b.jsx)(c,{id:`g`,title:`Mapping`,children:(0,b.jsx)(`p`,{className:`note`,children:`Nothing to do to it.`})}),"group/sub":()=>(0,b.jsx)(c,{id:`g`,title:`Job 1`,actions:S(`Rename`,`Remove`),children:(0,b.jsx)(c,{id:`g-sub`,title:`Scenarios`,level:3,actions:S(`Clear`),children:(0,b.jsx)(`p`,{className:`note`,children:`Inside a pane.`})})}),"group/depth":()=>(0,b.jsxs)(c,{id:`g`,title:`Job 1`,actions:S(`Rename`,`Remove`),children:[(0,b.jsxs)(c,{id:`g-scenarios`,title:`Scenarios`,level:3,children:[(0,b.jsx)(c,{id:`g-plain`,title:`Plain`,level:3,actions:S(`Rename`,`Remove`),children:(0,b.jsx)(`p`,{className:`note`,children:`A pane, a section of it, and a thing with a name inside that section.`})}),(0,b.jsx)(c,{id:`g-strict`,title:`Strict`,level:3,actions:S(`Rename`,`Remove`)})]}),(0,b.jsx)(c,{id:`g-output`,title:`Output`,level:3,children:(0,b.jsx)(`p`,{className:`note`,children:`The next section of the same pane.`})})]}),"group/renaming":()=>(0,b.jsx)(E,{}),"menu/disabled":()=>(0,b.jsx)(c,{id:`g`,title:`Scenario 4`,actions:[{label:`Rename`,icon:p.rename,onSelect:x},{label:`Clone`,icon:p.clone,onSelect:x,disabled:!0},{label:`Remove`,icon:p.remove,onSelect:x,destructive:!0}]}),"repeater/empty":()=>(0,b.jsx)(C,{start:[]}),"repeater/min":()=>(0,b.jsx)(C,{start:[`Scenario 1`],min:1}),"repeater/between":()=>(0,b.jsx)(C,{start:[`Scenario 1`,`Scenario 2`],min:1,max:4}),"repeater/max":()=>(0,b.jsx)(C,{start:[`Scenario 1`,`Scenario 2`,`Scenario 3`,`Scenario 4`],min:1,max:4}),"props/controls":()=>(0,b.jsx)(s,{label:`Controls`,onRename:x,onRemove:x,onAdd:x,addLabel:`Add row`,rows:[j,{key:`language`,label:`{language}`,code:!0,control:{type:`text`,value:`English UK`,onChange:x}},{key:`geo`,label:`{geo}…{/geo}`,code:!0,control:{type:`static`,text:`Block`},enabled:{checked:!0,onChange:x}},{key:`image`,label:`Image`,control:{type:`checkbox`,checked:!0,label:`Send`,onChange:x}}]}),"props/required":()=>(0,b.jsx)(s,{label:`Output`,onRemove:x,rows:[j]}),"props/long":()=>(0,b.jsx)(s,{label:`Long`,onRemove:x,rows:[{key:`geo-location`,label:`{geo-location}`,code:!0,control:{type:`text`,value:`Hamilton, Waikato, New Zealand, Southern Hemisphere, Earth`,onChange:x}},{key:`averylongtokennamethatdoesnotbreak`,label:`{averylongtokennamethatdoesnotbreak}`,code:!0,control:{type:`text`,value:``,onChange:x}}]}),"props/widths":()=>(0,b.jsx)(s,{label:`Widths`,rows:[{key:`type`,label:`Type`,required:!0,control:{type:`select`,value:`ollama`,onChange:x,options:[{value:`ollama`,label:`Ollama (Local)`},{value:`anthropic`,label:`Anthropic`}]}},{key:`url`,label:`Address`,control:{type:`text`,value:``,placeholder:`the lab's Ollama`,onChange:x}},{key:`model`,label:`Model`,control:{type:`text`,value:`minicpm-v4.6:latest`,onChange:x}},{key:`seed`,label:`Seed`,control:{type:`text`,value:``,placeholder:`1234`,marker:`default`,onChange:x}},{key:`nPredict`,label:`Reply tokens`,control:{type:`text`,value:``,size:`num`,marker:`server`,onChange:x}},{key:`temperature`,label:`Temperature`,control:{type:`text`,value:`0.2`,onChange:x}},{key:`grammar`,label:`Grammar`,control:{type:`textarea`,value:`root ::= item (", " item)*
-item ::= [a-z ]+`,onChange:x}}]}),"props/secret":()=>(0,b.jsx)(s,{label:`Keys`,rows:[{key:`held`,label:`Key`,control:{type:`secret`,value:`sk-not-a-real-key-0000`,onChange:x}},{key:`empty`,label:`Key`,control:{type:`secret`,value:``,onChange:x}}]}),"props/code":()=>(0,b.jsx)(s,{label:`Debug`,rows:[{key:`endpoint`,label:`Endpoint`,required:!0,control:{type:`code`,text:`https://api.anthropic.com/v1/messages`}},{key:`prompt`,label:`Prompt`,required:!0,control:{type:`static`,text:`(prompt)`}},{key:`model`,label:`model`,control:{type:`code`,text:`claude-sonnet-4-5-20250929`}},{key:`options`,label:`options`,control:{type:`code`,text:`{"seed":1234,"temperature":0}`}},{key:`stream`,label:`stream`,control:{type:`code`,text:`false`}}]}),"table/sortable":()=>(0,b.jsx)(I,{count:4}),"table/paged":()=>(0,b.jsx)(I,{count:35,pageSize:10}),"table/filter":()=>(0,b.jsx)(F,{}),"table/filter-on":()=>(0,b.jsx)(F,{search:`IMG_40`,size:`large`}),"edit/form":()=>(0,b.jsx)(D,{outcome:`Connected`,menu:[{label:`Restore…`,icon:p.restore,onSelect:x},{label:`Apply preset: repeat 1.3`,icon:p.preset,onSelect:x}],sections:[{key:`identity`,rows:O()},{key:`settings`,title:`Settings`,rows:k()},{key:`debug`,title:`Debug`,rows:A(),folded:!0}]}),"edit/plain":()=>(0,b.jsx)(o,{title:`A mapping`,sections:[{key:`one`,rows:[j,{key:`with`,label:`Separator`,control:{type:`text`,value:`, `,onChange:x}}]}],onClose:x}),"edit/long":()=>(0,b.jsx)(D,{menu:[{label:`Restore…`,icon:p.restore,onSelect:x,disabled:!0}],sections:[{key:`identity`,rows:O()},{key:`settings`,title:`Settings`,rows:[...k(),{key:`grammar`,label:`Grammar`,control:{type:`textarea`,value:`root ::= item (", " item)*
-item ::= [a-z ]+`,onChange:x}},{key:`imagePosition`,label:`Image position`,control:{type:`select`,value:`after`,onChange:x,options:[{value:`before`,label:`before the prompt`},{value:`after`,label:`after the prompt`}]}},{key:`cachePrompt`,label:`Cache prompt`,control:{type:`select`,value:`off`,onChange:x,options:[{value:`on`,label:`on`},{value:`off`,label:`off`}]}}]},{key:`debug`,title:`Debug`,rows:A(),folded:!0}]}),"edit/status":()=>(0,b.jsx)(D,{status:`the relay could not reach http://127.0.0.1:11434`,statusBad:!0,sections:[{key:`identity`,rows:O()}]}),"view/tokens":()=>(0,b.jsx)(c,{id:`g`,title:`Token mappings`,level:3,children:(0,b.jsx)(w,{})}),"view/output":()=>(0,b.jsx)(T,{start:{kind:`list`,parse:`csv`,modifiers:[{type:`case`,mode:`LOWER`}]}}),"view/output-text":()=>(0,b.jsx)(T,{start:{kind:`text`,modifiers:[]}}),"view/output-two":()=>(0,b.jsx)(T,{start:{kind:`list`,parse:`csv`,modifiers:[{type:`case`,mode:`TITLE`},{type:`join`,with:` | `}]}}),"view/output-twice":()=>(0,b.jsx)(T,{start:{kind:`list`,parse:`csv`,modifiers:[{type:`case`,mode:`TITLE`},{type:`case`,mode:`LOWER`}]}})},N=y.map(e=>({...e,render:M[e.path]})),P=e=>Array.from({length:e},(e,t)=>({name:`IMG_${String(4e3+t*7919%97).padStart(4,`0`)}.jpg`,kb:t*37%900+40}));function F({search:e=``,size:t=`all`}){let[n,r]=(0,m.useState)(e),[i,o]=(0,m.useState)(t),[s,l]=(0,m.useState)(`shoot`),d=P(12),f=d.filter(e=>e.name.toLowerCase().includes(n.trim().toLowerCase())&&(i===`all`||i===`large`==e.kb>=400));return(0,b.jsxs)(c,{id:`files`,title:`Files`,children:[(0,b.jsx)(u,{noun:`files`,search:{label:`Name`,placeholder:`Search names`,value:n,onChange:r},count:{shown:f.length,total:d.length},filters:[{key:`source`,label:`Source`,value:s,neutral:null,onChange:l,options:[{value:`shoot`,label:`Tuesday shoot`},{value:`samples`,label:`Sample library`}]},{key:`size`,label:`Size`,value:i,onChange:o,options:[{value:`all`,label:`All`},{value:`small`,label:`Small`},{value:`large`,label:`Large`}]}]}),f.length?(0,b.jsx)(a,{label:`Files`,rows:f,rowKey:e=>e.name,columns:[{key:`name`,header:`Name`,cell:e=>e.name},{key:`size`,header:`Size`,sort:e=>e.kb,cell:e=>`${e.kb} KB`}]}):(0,b.jsx)(`p`,{className:`note`,children:`No files match these filters.`})]})}function I({count:e,pageSize:t}){let n=P(e);return(0,b.jsx)(c,{id:`files`,title:`Files`,children:(0,b.jsx)(a,{label:`Files`,rows:n,rowKey:e=>e.name,pageSize:t??null,columns:[{key:`name`,header:`Name`,cell:e=>e.name},{key:`size`,header:`Size`,sort:e=>e.kb,cell:e=>`${e.kb} KB`},{key:`menu`,header:``,sort:null,cell:e=>(0,b.jsx)(i,{label:`${e.name} actions`,actions:[{label:`Rename`,icon:p.rename,onSelect:x},{label:`Remove`,icon:p.remove,destructive:!0,onSelect:x}]})}]})})}var L=()=>decodeURIComponent(location.hash.replace(/^#\/?/,``));function R(){let[e,t]=(0,m.useState)(L);(0,m.useEffect)(()=>{let e=()=>t(L());return window.addEventListener(`hashchange`,e),()=>window.removeEventListener(`hashchange`,e)},[]);let n=N.find(t=>t.path===e);return n?(0,b.jsxs)(`main`,{className:`gallery-one`,"data-route":n.path,children:[(0,b.jsxs)(`nav`,{className:`gallery-nav`,children:[(0,b.jsx)(`a`,{href:`#/`,children:`All`}),(0,b.jsx)(`span`,{children:n.title})]}),n.render()]}):(0,b.jsxs)(`main`,{className:`gallery-index`,"data-route":``,children:[(0,b.jsx)(`h1`,{children:`Components`}),(0,b.jsx)(`ul`,{children:N.map(e=>(0,b.jsxs)(`li`,{children:[(0,b.jsx)(`a`,{href:`#/${e.path}`,children:e.title}),` `,(0,b.jsx)(`code`,{children:e.path})]},e.path))})]})}(0,h.createRoot)(document.getElementById(`root`)).render((0,b.jsx)(m.StrictMode,{children:(0,b.jsx)(R,{})}));