Make three structural failures explicit and distinguish direct API completion from cancelled Batch attempt
007dd84 verified | <html lang="zh-CN"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1"><title>VideoGLaMM · ActivityNet Object Review · Luna</title><link rel="stylesheet" href="style.css"></head> | |
| <body><header><div><h1>VideoGLaMM · ActivityNet Object Review · Luna</h1><div class="sub">gpt-6-luna · 直接 Responses API · 相同 20 个样本 / 完全相同 prompt</div></div><div class="toolbar"><button id="prev">← 上一个</button><button id="next">下一个 →</button><a class="baseline" href="https://huggingface.co/spaces/Viviannnyang/VideoGLaMM_ActivityNet_Object_Review20" target="_blank">上一轮 Sol ↗</a><button id="export">导出审核</button><button id="import">导入审核</button><input id="import-file" type="file" accept="application/json" hidden></div></header> | |
| <div class="notice"><strong>模型候选，尚待人工审核。</strong>109 个实体组 / 187 处 text spans；23 个 mask 对应候选，86 个待补或待审；完整对应候选 0 / 20。全部实体 text 均保留。结构校验 17/20 通过，3 条待修正。<span id="processing"></span></div> | |
| <main><aside class="sidebar"><div class="sidebar-head"><strong id="list-count">20 samples</strong><input id="search" aria-label="搜索样本或 caption" placeholder="搜索 ID / caption / 实体"><select id="filter" aria-label="筛选审核状态"><option value="all">全部样本</option><option value="unreviewed">尚未审核</option><option value="reviewed">已审核并记录</option><option value="needs_review">需修改</option><option value="reject">不采用</option></select><small id="review-count"></small></div><nav id="entry-list" aria-label="样本列表"></nav></aside> | |
| <section class="content"><div id="loading">正在载入…</div><article id="entry" hidden> | |
| <div class="entry-head"><div><h2 id="title"></h2><div id="meta" class="sub"></div></div><div id="badges" class="badges"></div></div> | |
| <section class="panel"><div class="section-head"><h3>Caption · 点击高亮 text 选择实体组</h3><button id="all-groups">显示全部高亮</button></div><div id="caption" class="caption"></div><div id="selected-summary" class="sub"></div></section> | |
| <div class="grid"><section class="panel"><div class="section-head"><h3>采样帧与源 mask</h3><button id="reference">回到 mask 参考帧</button></div><div class="media-tools"><label>显示 <select id="mode"><option value="all">全部源 mask</option><option value="rgb">原始 RGB</option><option value="selected">单个源 mask</option><option value="binary">二值 mask</option></select></label><label>源 mask <select id="mask" aria-label="选择源 mask"></select></label><button id="play">播放采样帧</button></div><div id="frame-state" class="frame-state"></div><div class="big-frame"><img id="big-image" alt="当前采样帧"></div><div id="frames" class="frames" aria-label="十张采样帧"></div><div id="timing" class="sub"></div><p class="sub">只有参考帧有现成 mask / bbox。其余帧显示 RGB，不传播 mask。播放为 2 张/秒的检查用幻灯片；原视频 FPS 未提供。</p></section> | |
| <section class="panel inventory-panel"><h3>原始 mask 清单</h3><p class="sub">源类别名仅作线索，可能错标。点击检查该 mask；不会更改实体的候选对应。</p><div id="objects"></div><hr><h3>当前实体 → text / mask</h3><div id="group-detail"></div><div id="mask-preview"></div></section></div> | |
| <section class="panel"><div class="section-head"><h3>Caption-to-object groups · 含缺 mask 实体</h3><select id="group-filter" aria-label="筛选实体组"><option value="all">全部实体组</option><option value="matched">有对应候选</option><option value="pending">待补 / 待审</option></select></div><div id="groups" class="groups"></div></section> | |
| <section class="panel"><h3>问题与整理记录</h3><div id="issues"></div></section> | |
| <section class="panel review-panel"><h3>审核记录</h3><p class="sub">审核状态与备注仅保存在当前浏览器，可导出 / 导入；不回写源标注。标为“已审核”也不代表缺 mask 已补齐。</p><label for="review-status">审核状态</label> <select id="review-status"><option value="unreviewed">尚未审核</option><option value="reviewed">已审核并记录</option><option value="needs_review">需修改</option><option value="reject">不采用</option></select><label for="note">审核备注</label><textarea id="note" placeholder="例如 g2 text 范围、指代问题，m0 区域是否正确，待补 mask…"></textarea><div id="saved" role="status"></div></section> | |
| <details class="panel"><summary>原始 caption / 完整记录 / GPT 实际看到的 mask 证据</summary><h3>原始 refined_caption</h3><pre id="original-refined"></pre><h3>原始 new_caption</h3><pre id="original-new"></pre><div id="links" class="download-links"></div><div class="evidence"><figure><img id="gpt-overlay" alt="GPT 输入：全部源 mask 带 ID 叠加图"><figcaption>GPT 输入：全部 mask 带 ID 叠加图</figcaption></figure><figure><img id="gpt-identities" alt="GPT 输入：身份裁剪图"><figcaption>GPT 输入：身份裁剪图</figcaption></figure><figure><img id="gpt-binary" alt="GPT 输入：二值 mask 图"><figcaption>GPT 输入：原始二值 mask 图</figcaption></figure></div><p class="sub">每条请求另含全部 10 张 clean RGB，共 13 个 image input blocks。</p><pre id="raw-json"></pre></details> | |
| </article></section></main><footer><a href="data/annotations.json" download>全部 20 条标注 JSON</a><a href="data/annotation_prompt.txt" target="_blank">完整实际 prompt</a><a href="data/TEXT_ANNOTATION.original.md" target="_blank">原样标注规则</a><a href="data/prompt_audit.md" target="_blank">规则核对</a><a href="data/result_statistics.json" target="_blank">Luna 统计</a><a href="data/request_audit.json" target="_blank">输入一致性核对</a><a href="data/model_comparison.json" target="_blank">两轮统计（非准确率）</a><a href="data/processing_receipt.json" target="_blank">处理完成凭据</a><a href="https://github.com/mbzuai-oryx/VideoGLaMM/blob/main/Dataset.md" target="_blank">数据来源</a></footer><script src="app.js"></script></body></html> | |