trajectory-reader / index.html
MagicCaster's picture
trajectory reader: page
e640518 verified
Raw History Blame Contribute Delete
35.6 kB
<!doctype html>
<html lang="zh-CN">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>Trajectory Reader</title>
<!--
Trajectory Reader — render the `messages` list of any Hugging Face dataset row.
Deployed as the static Space NewsBreak/trajectory-reader. No server, no bundled
data, no bundled token, and no token box: viewers sign in with their own Hugging
Face account (OAuth PKCE, scope read-repos, restricted to NewsBreak members) and
every request goes straight from their browser to huggingface.co carrying THEIR
access token. So the Hub's own permission check applies: someone without access
to a private dataset gets 401/403 and sees nothing. Nobody ever types a
credential into this page.
Data path: datasets-server.huggingface.co /splits + /rows (paged, no download).
Fallback when the dataset isn't processed by the viewer backend: stream the
raw .jsonl via /resolve and split lines client-side.
-->
<link rel="preconnect" href="https://fonts.googleapis.com">
<link rel="stylesheet" href="https://fonts.googleapis.com/css2?family=IBM+Plex+Sans:wght@400;500;600&family=IBM+Plex+Mono:wght@400;500&display=swap">
<style>
:root{
--bg:#F6F5F1; --panel:#FFFFFF; --ink:#1B1A17; --muted:#6E6A60; --line:#DDD9CF; --line-soft:#ECE9E1;
--accent:#0F8A78; --accent-ink:#0B6B5D; --danger:#B4373A;
--user:#3D5AD6; --assistant:#0F8A78; --tool:#B0700E; --system:#7A766C; --reason:#7A5AB8;
--user-bg:#EEF1FC; --assistant-bg:#EAF5F2; --tool-bg:#FBF3E4; --system-bg:#F0EFEA; --reason-bg:#F3EFFA;
--code-bg:#F1EFE9; --shadow:0 1px 2px rgba(27,26,23,.06);
--sans:"IBM Plex Sans",system-ui,-apple-system,"Segoe UI",sans-serif;
--mono:"IBM Plex Mono",ui-monospace,SFMono-Regular,Menlo,Consolas,monospace;
}
@media (prefers-color-scheme: dark){
:root:not([data-theme="light"]){
--bg:#17181C; --panel:#1F2126; --ink:#ECEAE3; --muted:#9A968C; --line:#2E3038; --line-soft:#272930;
--accent:#3FB8A4; --accent-ink:#6FD3C2; --danger:#E06B6E;
--user:#8C9DF0; --assistant:#4FC2AD; --tool:#E0A54A; --system:#A5A197; --reason:#B39BE6;
--user-bg:#20253A; --assistant-bg:#1B2E2B; --tool-bg:#302819; --system-bg:#24262B; --reason-bg:#2A2436;
--code-bg:#15161A; --shadow:none;
}
}
:root[data-theme="dark"]{
--bg:#17181C; --panel:#1F2126; --ink:#ECEAE3; --muted:#9A968C; --line:#2E3038; --line-soft:#272930;
--accent:#3FB8A4; --accent-ink:#6FD3C2; --danger:#E06B6E;
--user:#8C9DF0; --assistant:#4FC2AD; --tool:#E0A54A; --system:#A5A197; --reason:#B39BE6;
--user-bg:#20253A; --assistant-bg:#1B2E2B; --tool-bg:#302819; --system-bg:#24262B; --reason-bg:#2A2436;
--code-bg:#15161A; --shadow:none;
}
*{box-sizing:border-box}
[hidden]{display:none!important}
html,body{margin:0;background:var(--bg);color:var(--ink);font:14px/1.5 var(--sans)}
button,input,select{font:inherit;color:inherit}
a{color:var(--accent-ink)}
:focus-visible{outline:2px solid var(--accent);outline-offset:2px}
@media (prefers-reduced-motion: reduce){*{transition:none!important;animation:none!important}}
/* ---------- toolbar ---------- */
.bar{position:sticky;top:0;z-index:5;background:var(--panel);border-bottom:1px solid var(--line);padding-block:10px;padding-inline:20px}
.bar-row{display:flex;flex-wrap:wrap;align-items:center;gap:8px 12px;max-width:1400px;margin:0 auto}
.brand{font-weight:600;letter-spacing:.01em;margin-right:6px;white-space:nowrap}
.brand small{color:var(--muted);font-weight:400;margin-left:6px}
.field{display:flex;align-items:center;gap:6px}
.field label{color:var(--muted);font-size:12px;text-transform:uppercase;letter-spacing:.06em;white-space:nowrap}
input[type=text],input[type=password],input[type=number],select{
background:var(--bg);border:1px solid var(--line);border-radius:6px;padding:6px 9px;min-width:0}
input[type=text]{width:300px;max-width:100%}
input[type=password]{width:210px}
input[type=number]{width:84px;font-variant-numeric:tabular-nums}
select{max-width:260px}
.btn{background:var(--panel);border:1px solid var(--line);border-radius:6px;padding:6px 11px;cursor:pointer;white-space:nowrap}
.btn:hover{border-color:var(--accent)}
.btn.primary{background:var(--accent);border-color:var(--accent);color:#fff}
.btn.primary:hover{filter:brightness(1.05)}
.btn.quiet{border-color:transparent;color:var(--muted)}
.btn:disabled{opacity:.45;cursor:default}
a.btn{text-decoration:none;display:inline-block}
.btn.hf{background:#FFD21E;border-color:#E8BE10;color:#1B1A17;font-weight:600}
.btn.hf::before{content:"🤗";margin-right:6px}
.who-name{color:var(--muted);font-size:13px;white-space:nowrap}
.who-name b{color:var(--ink);font-weight:500}
.nav{display:flex;align-items:center;gap:6px;margin-left:auto}
.nav .count{color:var(--muted);font-variant-numeric:tabular-nums;white-space:nowrap}
.status{max-width:1400px;margin:8px auto 0;font-size:12.5px;color:var(--muted);min-height:1.2em;display:flex;gap:10px;align-items:center;flex-wrap:wrap}
.status.err{color:var(--danger)}
.progress{height:3px;background:var(--line-soft);border-radius:2px;flex:1 1 160px;overflow:hidden}
.progress i{display:block;height:100%;width:0;background:var(--accent);transition:width .2s}
/* ---------- layout ---------- */
.wrap{max-width:1400px;margin:0 auto;padding-block:18px;padding-inline:20px;display:grid;grid-template-columns:300px minmax(0,1fr);gap:22px}
@media (max-width:900px){.wrap{grid-template-columns:1fr}}
.rail{display:flex;flex-direction:column;gap:14px;align-self:start;position:sticky;top:74px;max-height:calc(100vh - 90px);overflow:auto}
@media (max-width:900px){.rail{position:static;max-height:none}}
.card{background:var(--panel);border:1px solid var(--line);border-radius:8px;padding:12px 14px;box-shadow:var(--shadow)}
.card h3{margin:0 0 8px;font-size:12px;text-transform:uppercase;letter-spacing:.07em;color:var(--muted);font-weight:600}
.kv{display:grid;grid-template-columns:auto minmax(0,1fr);gap:4px 10px;font-size:13px}
.kv dt{color:var(--muted);white-space:nowrap}
.kv dd{margin:0;overflow-wrap:anywhere}
.kv dd.num{font-variant-numeric:tabular-nums}
.opts{display:flex;flex-direction:column;gap:6px;font-size:13px}
.opts label{display:flex;align-items:center;gap:8px;cursor:pointer}
details{border-top:1px solid var(--line-soft)}
details:first-of-type{border-top:0}
summary{cursor:pointer;padding:5px 0;color:var(--ink);list-style:none;display:flex;gap:8px;align-items:baseline}
summary::-webkit-details-marker{display:none}
summary::before{content:"▸";color:var(--muted);font-size:11px;flex:none}
details[open]>summary::before{content:"▾"}
summary .hint{color:var(--muted);font-size:12px;margin-left:auto;white-space:nowrap}
pre{margin:0;font:12.5px/1.5 var(--mono);white-space:pre-wrap;overflow-wrap:anywhere;background:var(--code-bg);border-radius:6px;padding:8px 10px;max-height:420px;overflow:auto}
.rail pre{max-height:260px}
/* ---------- transcript ---------- */
.main{min-width:0;max-width:880px}
.empty{color:var(--muted);padding:40px 0;text-align:center}
.turn{display:grid;grid-template-columns:82px minmax(0,1fr);gap:12px;padding:12px 0;border-top:1px solid var(--line-soft)}
.turn:first-child{border-top:0}
.who{display:flex;flex-direction:column;gap:3px;align-items:flex-start}
.pill{font-size:11px;font-weight:600;letter-spacing:.06em;text-transform:uppercase;padding:2px 8px;border-radius:999px;color:#fff;background:var(--system)}
.turn[data-role=user] .pill{background:var(--user)}
.turn[data-role=assistant] .pill{background:var(--assistant)}
.turn[data-role=tool] .pill{background:var(--tool)}
.who .n{color:var(--muted);font-size:11px;font-variant-numeric:tabular-nums}
.body{display:flex;flex-direction:column;gap:8px;min-width:0}
.bubble{border-radius:8px;padding:10px 12px;background:var(--system-bg);border:1px solid transparent}
.turn[data-role=user] .bubble.text{background:var(--user-bg)}
.turn[data-role=assistant] .bubble.text{background:var(--assistant-bg)}
.turn[data-role=tool] .bubble.text{background:var(--tool-bg)}
.text .t{white-space:pre-wrap;overflow-wrap:anywhere}
.reason{background:var(--reason-bg);border-left:3px solid var(--reason)}
.reason summary{color:var(--reason);font-weight:500}
.reason .t{white-space:pre-wrap;overflow-wrap:anywhere;color:var(--ink);opacity:.9;padding:4px 0 2px}
.call{background:var(--tool-bg);border-left:3px solid var(--tool)}
.call .head{display:flex;gap:8px;align-items:baseline;flex-wrap:wrap;font-family:var(--mono);font-size:12.5px}
.call .head b{color:var(--tool);font-weight:600}
.call .head .id{color:var(--muted);font-size:11px;margin-left:auto}
.call pre{margin-top:6px;background:transparent;padding:0;max-height:360px}
.call.bad{border-left-color:var(--danger)}
.call.bad .head b{color:var(--danger)}
.call .warn{margin-top:6px;font-size:12px;color:var(--danger)}
.meta{font-family:var(--mono);font-size:11.5px;color:var(--muted);display:flex;gap:10px;flex-wrap:wrap}
.more{margin-top:6px}
.more .btn{font-size:12px;padding:3px 9px}
.hidden-note{color:var(--muted);font-size:12px;padding:6px 0;border-top:1px dashed var(--line)}
.section-h{font-size:12px;text-transform:uppercase;letter-spacing:.07em;color:var(--muted);font-weight:600;margin:14px 0 4px}
.rawrow pre{max-height:70vh}
kbd{font:11px var(--mono);background:var(--code-bg);border:1px solid var(--line);border-bottom-width:2px;border-radius:4px;padding:0 5px}
.help{color:var(--muted);font-size:12px}
.help kbd{margin:0 2px}
.tools-list{display:flex;flex-direction:column}
.tools-list summary b{font-family:var(--mono);font-weight:500}
.tools-list .d{color:var(--muted);font-size:12px;padding-bottom:6px}
</style>
</head>
<body>
<div class="bar">
<div class="bar-row">
<div class="brand">Trajectory Reader<small>HF messages</small></div>
<div class="field"><label for="repo">dataset</label><input id="repo" type="text" placeholder="org/name" spellcheck="false"></div>
<div class="field" id="authOauth" hidden>
<a id="signin" class="btn hf" href="#" target="_blank" rel="noopener" hidden>Sign in with Hugging Face</a>
<span id="who" class="who-name" hidden></span>
<button id="signout" class="btn quiet" hidden>退出</button>
</div>
<button id="load" class="btn primary">加载</button>
<div class="field"><label for="split">split</label><select id="split" disabled><option>—</option></select></div>
<div class="nav">
<button id="prev" class="btn" disabled title="上一行 (←)">‹</button>
<input id="idx" type="number" min="0" value="0" disabled>
<span class="count" id="count">/ —</span>
<button id="next" class="btn" disabled title="下一行 (→)">›</button>
<button id="rand" class="btn" disabled title="随机 (r)">随机</button>
</div>
</div>
<div class="status" id="status"></div>
</div>
<div class="wrap">
<aside class="rail">
<div class="card">
<h3>行信息</h3>
<dl class="kv" id="rowmeta"><dt>状态</dt><dd>尚未加载</dd></dl>
</div>
<div class="card">
<h3>显示</h3>
<div class="opts">
<label><input type="checkbox" id="optReason" checked> 默认展开 reasoning</label>
<label><input type="checkbox" id="optTools"> 隐藏 tool 返回</label>
<label><input type="checkbox" id="optSystem"> 隐藏 system</label>
<label><input type="checkbox" id="optRaw"> 显示整行原始 JSON</label>
</div>
<p class="help" style="margin:10px 0 0"><kbd>←</kbd><kbd>→</kbd> 翻行 · <kbd>r</kbd> 随机</p>
</div>
<div class="card" id="toolsCard" hidden>
<h3>tools <span id="toolsN"></span></h3>
<div class="tools-list" id="toolsList"></div>
</div>
<div class="card" id="colsCard" hidden>
<h3>其它列</h3>
<div id="cols"></div>
</div>
</aside>
<main class="main" id="main">
<div class="empty" id="intro"></div>
</main>
</div>
<script>
(() => {
'use strict';
const DS = 'https://datasets-server.huggingface.co';
const HUB = 'https://huggingface.co';
const PAGE = 20; // rows per /rows request (server max 100)
const CLIP = 1800; // chars shown before "展开" on long blocks
const LS_REPO = 'trajreader.repo', LS_OAUTH = 'trajreader.oauth', SS_HASH = 'trajreader.pendingHash';
const HFJS = 'https://cdn.jsdelivr.net/npm/@huggingface/hub@2.17.1/+esm'; // OAuth PKCE helper (Space mode only)
const ORG_ID = '6a2a067c086133dad7d362e4'; // NewsBreak — preselects the org on the consent screen
const $ = id => document.getElementById(id);
const el = {
repo: $('repo'), load: $('load'), split: $('split'),
authOauth: $('authOauth'), signin: $('signin'), signout: $('signout'), who: $('who'), intro: $('intro'),
prev: $('prev'), next: $('next'), rand: $('rand'), idx: $('idx'), count: $('count'),
status: $('status'), rowmeta: $('rowmeta'), main: $('main'), cols: $('cols'), colsCard: $('colsCard'),
toolsCard: $('toolsCard'), toolsList: $('toolsList'), toolsN: $('toolsN'),
optReason: $('optReason'), optTools: $('optTools'), optSystem: $('optSystem'), optRaw: $('optRaw'),
};
const S = {
repo: '', splits: [], sel: null, total: null, idx: 0,
mode: 'server', // 'server' (datasets-server /rows) | 'raw' (stream .jsonl)
cache: new Map(), // row_idx -> row object
rawLines: null, // raw mode: array of parsed rows
files: [], // raw mode: candidate files
current: null, loadingRow: null, features: [], gen: 0,
};
/* ---------- storage (this browser only; never a credential typed by hand) ---------- */
const ls = {
get(k){ try { return localStorage.getItem(k) || ''; } catch { return ''; } },
set(k,v){ try { v ? localStorage.setItem(k,v) : localStorage.removeItem(k); } catch {} },
};
el.repo.value = ls.get(LS_REPO) || 'NewsBreak/scout_c6_v2_norm';
/* ---------- auth: Hugging Face sign-in, and nothing else ----------
No token box on purpose: every request carries an OAuth token this page obtained
for the person in front of it, so access is exactly their Hub access and no
credential is ever typed into, or stored by, this page. */
const auth = { accessToken: '', user: '' };
const bearer = () => auth.accessToken;
async function initAuth(){
const vars = (window.huggingface && window.huggingface.variables) || null;
if (!vars || !vars.OAUTH_CLIENT_ID) {
// Opened outside the Space (a local copy of the file): sign-in is impossible here.
el.intro.innerHTML = '请从 <a href="https://huggingface.co/spaces/NewsBreak/trajectory-reader">Space</a> 打开';
el.load.disabled = true;
return;
}
el.authOauth.hidden = false;
el.intro.textContent = 'Login first';
let hub;
try { hub = await import(HFJS); }
catch (e) { setStatus('登录组件加载失败,请刷新页面重试:' + e.message, true); return; }
let res = null;
try { res = await hub.oauthHandleRedirectIfPresent(); }
catch (e) { setStatus('登录失败:' + e.message, true); }
if (res) {
ls.set(LS_OAUTH, JSON.stringify(res));
let pending = ''; try { pending = sessionStorage.getItem(SS_HASH) || ''; sessionStorage.removeItem(SS_HASH); } catch {}
history.replaceState(null, '', location.pathname + pending);
} else {
try { const saved = JSON.parse(ls.get(LS_OAUTH) || 'null'); if (saved && new Date(saved.accessTokenExpiresAt).getTime() > Date.now() + 60e3) res = saved; } catch {}
}
if (res) {
auth.accessToken = res.accessToken;
auth.user = (res.userInfo && (res.userInfo.preferred_username || res.userInfo.name)) || '';
el.signin.hidden = true; el.who.hidden = false; el.signout.hidden = false;
el.intro.textContent = '点「加载」开始';
el.who.innerHTML = `已登录 <b>${esc(auth.user || 'HF user')}</b>`;
} else {
// Build the authorize URL NOW and hang it on a real link. Computing it inside the
// click handler means awaiting a fetch first, which breaks the user-gesture chain
// and gets the new tab blocked as a popup when the Space runs inside its iframe.
const embedded = window.self !== window.top;
try {
const url = (await hub.oauthLoginUrl({ redirectUrl: location.origin + location.pathname, scopes: vars.OAUTH_SCOPES || 'openid profile read-repos' })) + '&orgIds=' + ORG_ID;
el.signin.href = url;
if (!embedded) el.signin.removeAttribute('target'); // top-level: navigate in place
el.signin.hidden = false;
el.signin.addEventListener('click', () => {
try { sessionStorage.setItem(SS_HASH, location.hash); } catch {}
if (embedded) setStatus('登录已在新标签页打开;登录完成后请在那个标签页里继续使用');
});
} catch (e) {
setStatus('无法发起 HF 登录:' + (e.message || e), true);
}
}
el.signout.onclick = () => { ls.set(LS_OAUTH, ''); auth.accessToken = ''; location.reload(); };
}
/* ---------- helpers ---------- */
const esc = s => String(s).replace(/[&<>"']/g, c => ({'&':'&amp;','<':'&lt;','>':'&gt;','"':'&quot;',"'":'&#39;'}[c]));
const fmtN = n => n == null ? '—' : Number(n).toLocaleString('en-US');
const tryJSON = v => { if (typeof v !== 'string') return v; const t=v.trim(); if (!(t.startsWith('{')||t.startsWith('['))) return v; try { return JSON.parse(t); } catch { return v; } };
const pretty = v => typeof v === 'string' ? v : JSON.stringify(v, null, 2);
const isObj = v => v && typeof v === 'object' && !Array.isArray(v);
function setStatus(msg, err=false, progress=null){
el.status.className = 'status' + (err ? ' err' : '');
el.status.innerHTML = esc(msg) + (progress != null ? `<span class="progress"><i style="width:${Math.round(progress*100)}%"></i></span>` : '');
}
class HttpError extends Error { constructor(status, body, url){ super(`HTTP ${status}`); this.status=status; this.body=body; this.url=url; } }
async function hfFetch(url, opts={}){
const h = Object.assign({}, opts.headers||{});
const tok = bearer();
if (tok) h.Authorization = 'Bearer ' + tok;
const r = await fetch(url, Object.assign({}, opts, {headers:h}));
if (!r.ok){ let b=''; try { b = await r.text(); } catch {} throw new HttpError(r.status, b, url); }
return r;
}
function explain(e){
if (e instanceof HttpError){
let detail=''; try { detail = JSON.parse(e.body).error || ''; } catch { detail = (e.body||'').slice(0,160); }
if (e.status===401) return '401:请先登录 Hugging Face(登录可能已过期,点「退出」再登录一次)。';
if (e.status===403) return '403:你的账号没有访问该数据集的权限(登录时要勾选授权 NewsBreak 组织)。';
if (e.status===404) return '404:数据集不存在,或 viewer 后端尚未处理。' + (detail?` (${detail})`:'');
return `${e.status}: ${detail || e.message}`;
}
if (e instanceof TypeError) return '网络/CORS 错误:' + e.message;
return e.message || String(e);
}
/* ---------- URL hash (shareable; carries no credential) ---------- */
function writeHash(){
if (!S.repo) return;
const p = new URLSearchParams({repo:S.repo});
if (S.sel) { p.set('config', S.sel.config); p.set('split', S.sel.split); }
if (S.mode==='raw' && S.sel) p.set('file', S.sel.file);
p.set('idx', S.idx);
history.replaceState(null, '', '#' + p.toString());
}
function readHash(){
const p = new URLSearchParams(location.hash.slice(1));
return { repo:p.get('repo'), config:p.get('config'), split:p.get('split'), file:p.get('file'), idx:parseInt(p.get('idx')||'0',10)||0 };
}
/* ---------- dataset / splits ---------- */
async function loadDataset(want){
const repo = el.repo.value.trim().replace(/^datasets\//,'');
if (!repo) { setStatus('请输入数据集名,如 NewsBreak/scout_c6_v2_norm', true); return; }
if (!auth.accessToken) { setStatus('请先点「Sign in with Hugging Face」登录', true); return; }
ls.set(LS_REPO, repo);
const gen = ++S.gen; // a newer 加载 supersedes this one
Object.assign(S, {repo, splits:[], sel:null, total:null, idx:0, cache:new Map(), rawLines:null, files:[], current:null});
el.split.disabled = true; el.split.innerHTML = '<option>…</option>';
setStatus(`读取 ${repo} 的 splits…`);
try {
const r = await hfFetch(`${DS}/splits?dataset=${encodeURIComponent(repo)}`);
const j = await r.json();
S.mode = 'server';
S.splits = (j.splits||[]).map(s => ({config:s.config, split:s.split, label: (j.splits.some(o=>o.config!==s.config) ? `${s.config} / ` : '') + s.split}));
if (!S.splits.length) throw new Error('viewer 后端返回了 0 个 split');
} catch (e) {
// Fallback: list repo files and stream .jsonl directly.
setStatus(`viewer 后端不可用(${explain(e)}),改为直接读取仓库文件…`);
try {
const r = await hfFetch(`${HUB}/api/datasets/${repo}/tree/main?recursive=true`);
const files = (await r.json()).filter(f => f.type==='file' && /\.jsonl?$/i.test(f.path));
if (!files.length) throw new Error('仓库里没有 .jsonl / .json 文件');
S.mode = 'raw';
S.splits = files.map(f => ({config:'raw', split:f.path.replace(/\.jsonl?$/i,''), file:f.path, size:f.size, label:`${f.path} (${(f.size/1e6).toFixed(1)} MB)`}));
} catch (e2) { if (gen===S.gen) setStatus(explain(e2), true); return; }
}
if (gen !== S.gen) return;
el.split.innerHTML = S.splits.map((s,i)=>`<option value="${i}">${esc(s.label)}</option>`).join('');
el.split.disabled = false;
let pick = 0;
if (want){
const i = S.splits.findIndex(s => (want.file && s.file===want.file) || (s.split===want.split && (!want.config || s.config===want.config)));
if (i>=0) pick = i;
}
el.split.value = String(pick);
setStatus(`${repo} · ${S.splits.length} 个 split · ${S.mode==='server'?'viewer API':'直接读文件'}`);
await selectSplit(pick, want ? want.idx : 0);
}
async function selectSplit(i, idx=0){
S.sel = S.splits[i]; S.cache = new Map(); S.rawLines = null; S.total = null; S.idx = idx;
el.count.textContent = '/ …';
for (const b of [el.prev, el.next, el.rand, el.idx]) b.disabled = false;
if (S.mode==='raw') await streamRaw(S.sel);
await showRow(idx);
}
/* ---------- rows: datasets-server ---------- */
async function fetchPage(offset){
const {config, split} = S.sel;
const url = `${DS}/rows?dataset=${encodeURIComponent(S.repo)}&config=${encodeURIComponent(config)}&split=${encodeURIComponent(split)}&offset=${offset}&length=${PAGE}`;
const j = await (await hfFetch(url)).json();
S.total = j.num_rows_total; S.features = j.features || [];
for (const r of j.rows) S.cache.set(r.row_idx, {row:r.row, truncated:r.truncated_cells||[]});
}
async function getRow(idx){
if (S.mode==='raw') return S.rawLines[idx] ? {row:S.rawLines[idx], truncated:[]} : null;
if (!S.cache.has(idx)) await fetchPage(Math.floor(idx/PAGE)*PAGE);
return S.cache.get(idx) || null;
}
/* ---------- rows: raw .jsonl streaming (fallback) ---------- */
async function streamRaw(sel){
const url = `${HUB}/datasets/${S.repo}/resolve/main/${sel.file}`;
setStatus(`下载 ${sel.file}…`, false, 0);
const r = await hfFetch(url);
const reader = r.body.getReader(); const dec = new TextDecoder();
const total = sel.size || Number(r.headers.get('content-length')) || 0;
let got = 0, buf = '', rows = [];
S.rawLines = rows;
for (;;) {
const {done, value} = await reader.read();
if (done) break;
got += value.length; buf += dec.decode(value, {stream:true});
let nl;
while ((nl = buf.indexOf('\n')) >= 0) {
const line = buf.slice(0, nl).trim(); buf = buf.slice(nl+1);
if (line) { try { rows.push(JSON.parse(line)); } catch { rows.push({_parse_error:line.slice(0,200)}); } }
}
S.total = rows.length; el.count.textContent = `/ ${fmtN(rows.length)}…`;
setStatus(`下载 ${sel.file}… ${(got/1e6).toFixed(1)} MB,${fmtN(rows.length)} 行`, false, total ? got/total : null);
}
const tail = buf.trim(); if (tail) { try { rows.push(JSON.parse(tail)); } catch {} }
S.total = rows.length;
setStatus(`${sel.file} · ${fmtN(rows.length)} 行(已全部载入浏览器内存)`);
}
/* ---------- conversation detection & normalisation ---------- */
const CONV_KEYS = ['messages','conversations','conversation','chat','turns','trajectory','dialog','dialogue'];
const FROM_MAP = {human:'user', user:'user', gpt:'assistant', assistant:'assistant', model:'assistant', bot:'assistant',
system:'system', developer:'system', tool:'tool', observation:'tool', toolresult:'tool', tool_result:'tool', function:'tool', function_call:'assistant'};
const isMsgList = v => Array.isArray(v) && v.length>0 && v.every(m => isObj(m) && (('role' in m) || ('from' in m)));
function findConversation(row){
const keys = [...CONV_KEYS.filter(k=>k in row), ...Object.keys(row).filter(k=>!CONV_KEYS.includes(k))];
for (const k of keys){ const v = tryJSON(row[k]); if (isMsgList(v)) return {key:k, msgs:v}; }
// prompt/completion, chosen/rejected style
const sections = [];
for (const k of ['prompt','completion','chosen','rejected']) { const v = tryJSON(row[k]); if (isMsgList(v)) sections.push({key:k, msgs:v}); }
if (sections.length) return {key:sections.map(s=>s.key).join(' + '), sections};
return null;
}
function normMsg(m){
const role = (m.role || FROM_MAP[String(m.from||'').toLowerCase()] || m.from || 'unknown').toLowerCase();
let content = m.content ?? m.value ?? m.text ?? '';
let reasoning = m.reasoning ?? m.reasoning_content ?? m.reasoningContent ?? m.thinking ?? m.thought ?? '';
let calls = m.tool_calls ?? m.toolCalls ?? m.function_call ?? null;
if (calls && !Array.isArray(calls)) calls = [calls];
calls = calls ? calls.slice() : [];
// Anthropic-style content parts
if (Array.isArray(content)){
const texts = [];
for (const p of content){
if (!isObj(p)) { texts.push(String(p)); continue; }
const t = p.type;
if (t==='text' || t==='output_text' || t==='input_text') texts.push(p.text ?? '');
else if (t==='thinking' || t==='reasoning') reasoning = (reasoning ? reasoning+'\n\n' : '') + (p.thinking ?? p.text ?? '');
else if (t==='tool_use' || t==='toolCall' || t==='tool_call' || t==='function_call') calls.push(p);
else if (t==='tool_result' || t==='toolResult') texts.push(typeof p.content==='string' ? p.content : pretty(p.content));
else if (t==='image' || t==='image_url') texts.push('[image]');
else texts.push(pretty(p));
}
content = texts.join('\n');
} else if (content != null && typeof content !== 'string') content = pretty(content);
if (!reasoning && m.thinking_blocks) reasoning = (Array.isArray(m.thinking_blocks)?m.thinking_blocks:[m.thinking_blocks]).map(b => isObj(b) ? (b.thinking ?? b.text ?? '') : String(b)).filter(Boolean).join('\n\n');
// <think>…</think> embedded in content
if (!reasoning && typeof content==='string'){
const mt = content.match(/^\s*<think>([\s\S]*?)<\/think>\s*/);
if (mt) { reasoning = mt[1].trim(); content = content.slice(mt[0].length); }
}
const toolCalls = calls.map(c => {
const f = c.function || {};
let args = c.arguments ?? f.arguments ?? c.args ?? c.input ?? c.parameters ?? null;
args = tryJSON(args);
const name = c.name || f.name || c.tool || '';
// An empty name / empty arguments means the trajectory lost the call when it was
// rendered upstream — flag it instead of drawing a blank row.
const empty = !name || args == null || (isObj(args) && !Object.keys(args).length);
return { id: c.id ?? c.tool_call_id ?? c.call_id ?? '', name, args, type: c.type, empty };
});
return { role, content: String(content ?? ''), reasoning: String(reasoning ?? ''), toolCalls,
toolCallId: m.tool_call_id ?? m.toolCallId ?? m.call_id ?? '', name: m.name ?? m.toolName ?? m.tool_name ?? '', raw: m };
}
/* ---------- rendering ---------- */
function clipBlock(text, cls='t'){
const s = String(text);
if (s.length <= CLIP) return `<div class="${cls}">${esc(s)}</div>`;
const id = 'c' + Math.random().toString(36).slice(2,9);
return `<div class="${cls}" id="${id}" data-full="${esc(s)}">${esc(s.slice(0,CLIP))}<span class="muted">…</span></div>
<div class="more"><button class="btn" data-expand="${id}">展开全部 (${fmtN(s.length)} 字符)</button></div>`;
}
function renderTurn(m, i, opts){
const parts = [];
if (m.reasoning) parts.push(`<details class="bubble reason" ${opts.reason?'open':''}><summary>reasoning<span class="hint">${fmtN(m.reasoning.length)} 字符</span></summary>${clipBlock(m.reasoning)}</details>`);
if (m.content.trim()) parts.push(`<div class="bubble text">${clipBlock(m.content)}</div>`);
for (const c of m.toolCalls){
const nm = c.name ? esc(c.name) : '<i>(空 name)</i>';
const warn = c.empty ? `<div class="warn">↑ 这条 tool call 在数据里就是空的(缺失 name / arguments),不是渲染问题</div>` : '';
parts.push(`<div class="bubble call${c.empty?' bad':''}"><div class="head"><span>→</span><b>${nm}</b>${c.id?`<span class="id">${esc(c.id)}</span>`:''}</div>${c.args!=null ? `<pre>${esc(pretty(c.args))}</pre>` : ''}${warn}</div>`);
}
const meta = [];
if (m.role==='tool' && (m.name||m.toolCallId)) meta.push(`${m.name?esc(m.name):''}${m.toolCallId?` ← ${esc(m.toolCallId)}`:''}`);
if (!parts.length) parts.push(`<div class="bubble text"><span class="help">(空消息)</span></div>`);
return `<section class="turn" data-role="${esc(m.role)}"><div class="who"><span class="pill">${esc(m.role)}</span><span class="n">#${i}</span></div>
<div class="body">${meta.length?`<div class="meta">${meta.join(' ')}</div>`:''}${parts.join('')}</div></section>`;
}
function renderMessages(msgs, opts){
let hidden = 0, out = [];
msgs.forEach((raw, i) => {
const m = normMsg(raw);
if ((opts.hideTools && m.role==='tool') || (opts.hideSystem && m.role==='system')) { hidden++; return; }
if (hidden) { out.push(`<div class="hidden-note">已隐藏 ${hidden} 条</div>`); hidden = 0; }
out.push(renderTurn(m, i, opts));
});
if (hidden) out.push(`<div class="hidden-note">已隐藏 ${hidden} 条</div>`);
return out.join('');
}
function renderRail(row, convKey, truncated){
// row meta
const conv = convKey ? tryJSON(row[convKey]) : null;
const n = Array.isArray(conv) ? conv.length : null;
const roles = Array.isArray(conv) ? conv.reduce((a,m)=>{ const r=normMsg(m).role; a[r]=(a[r]||0)+1; return a; },{}) : null;
const kv = [
['行', `<span class="num">${fmtN(S.idx)}</span> / ${fmtN(S.total)}`],
['split', esc(S.sel.label)],
['来源', S.mode==='server' ? 'viewer API(分页)' : '原始 .jsonl'],
];
if (convKey) kv.push(['messages 列', `<code>${esc(convKey)}</code>`]);
if (n != null) kv.push(['轮次', `${n}` + (roles ? ` <span class="help">(${Object.entries(roles).map(([r,c])=>`${r} ${c}`).join(', ')})</span>` : '')]);
if (truncated.length) kv.push(['⚠ 截断', `viewer API 截断了:${truncated.map(esc).join(', ')}`]);
el.rowmeta.innerHTML = kv.map(([k,v]) => `<dt>${k}</dt><dd>${v}</dd>`).join('');
// tools
const tools = tryJSON(row.tools);
if (Array.isArray(tools) && tools.length){
el.toolsCard.hidden = false; el.toolsN.textContent = `(${tools.length})`;
el.toolsList.innerHTML = tools.map(t => {
const f = t.function || t; const name = f.name || t.name || '(unnamed)'; const d = f.description || t.description || '';
return `<details><summary><b>${esc(name)}</b></summary>${d?`<div class="d">${esc(d)}</div>`:''}<pre>${esc(pretty(f.parameters || f.input_schema || t))}</pre></details>`;
}).join('');
} else el.toolsCard.hidden = true;
// other columns
const skip = new Set([convKey, 'tools']);
const cols = Object.keys(row).filter(k => !skip.has(k));
el.colsCard.hidden = !cols.length;
el.cols.innerHTML = cols.map(k => {
const v = row[k]; const p = tryJSON(v);
if (v == null) return `<details><summary>${esc(k)}<span class="hint">null</span></summary></details>`;
if (typeof p === 'object') return `<details><summary>${esc(k)}<span class="hint">${Array.isArray(p)?`array[${p.length}]`:'object'}</span></summary><pre>${esc(pretty(p))}</pre></details>`;
const s = String(v);
if (s.length <= 80 && !s.includes('\n')) return `<details><summary>${esc(k)}<span class="hint" style="max-width:55%;overflow:hidden;text-overflow:ellipsis">${esc(s)}</span></summary></details>`;
return `<details><summary>${esc(k)}<span class="hint">${fmtN(s.length)} 字符</span></summary><pre>${esc(s)}</pre></details>`;
}).join('');
}
function paint(){
const cur = S.current; if (!cur) return;
const {row, truncated} = cur;
const opts = { reason: el.optReason.checked, hideTools: el.optTools.checked, hideSystem: el.optSystem.checked };
const conv = findConversation(row);
renderRail(row, conv && !conv.sections ? conv.key : null, truncated);
let html = '';
if (el.optRaw.checked) html += `<div class="rawrow"><div class="section-h">原始行 JSON</div><pre>${esc(JSON.stringify(row, (k,v)=>tryJSON(v), 2))}</pre></div>`;
if (!conv) html += `<div class="empty">这一行没有可识别的 messages 列(找 role/from 的对象数组)。打开「显示整行原始 JSON」看结构。</div>`;
else if (conv.sections) html += conv.sections.map(s => `<div class="section-h">${esc(s.key)}</div>${renderMessages(s.msgs, opts)}`).join('');
else html += renderMessages(conv.msgs, opts);
el.main.innerHTML = html;
window.scrollTo({top:0});
}
async function showRow(idx){
if (S.total != null) idx = Math.max(0, Math.min(idx, S.total-1));
idx = Math.max(0, idx|0);
S.idx = idx; el.idx.value = idx; S.loadingRow = idx;
try {
const r = await getRow(idx);
if (S.loadingRow !== idx) return; // superseded by a newer navigation
if (!r) { setStatus(`第 ${idx} 行不存在`, true); return; }
S.current = r; el.count.textContent = `/ ${fmtN(S.total)}`;
el.prev.disabled = idx<=0; el.next.disabled = S.total!=null && idx>=S.total-1;
paint(); writeHash();
if (S.mode==='server') setStatus(`${S.repo} · ${S.sel.label} · 第 ${fmtN(idx)} 行`);
} catch (e) { setStatus(explain(e), true); }
}
/* ---------- events ---------- */
el.load.onclick = () => loadDataset(null);
el.repo.addEventListener('keydown', e => { if (e.key==='Enter') loadDataset(null); });
el.split.onchange = () => selectSplit(parseInt(el.split.value,10), 0);
el.prev.onclick = () => showRow(S.idx-1);
el.next.onclick = () => showRow(S.idx+1);
el.rand.onclick = () => { if (S.total) showRow(Math.floor(Math.random()*S.total)); };
el.idx.addEventListener('change', () => showRow(parseInt(el.idx.value,10)||0));
for (const o of [el.optReason, el.optTools, el.optSystem, el.optRaw]) o.addEventListener('change', paint);
el.main.addEventListener('click', e => {
const b = e.target.closest('[data-expand]'); if (!b) return;
const box = document.getElementById(b.dataset.expand); if (!box) return;
box.textContent = box.dataset.full; b.parentElement.remove();
});
document.addEventListener('keydown', e => {
if (e.target.matches('input,select,textarea') || !S.sel) return;
if (e.key==='ArrowLeft') { e.preventDefault(); el.prev.click(); }
else if (e.key==='ArrowRight') { e.preventDefault(); el.next.click(); }
else if (e.key==='r') el.rand.click();
});
/* ---------- boot: resolve auth, then restore from hash if we can already authenticate ---------- */
(async () => {
await initAuth();
const h = readHash();
if (h.repo) el.repo.value = h.repo;
if (h.repo && bearer()) loadDataset(h);
})();
})();
</script>
</body>
</html>