WeaveBench / tasks /DOC /DOC_task_2_heading_style_normalize.md
wanlilll's picture
Add 114 tasks under 8-domain flat layout
71e9dba verified
metadata
id: DOC_task_2_heading_style_normalize
name:  LibreOffice 15 个混合样式标题 (8 H1 + 2 H2 + 5 fake)
category: DOC
timeout_seconds: 1800

Prompt

/tmp_workspace/report.odt contains 15 visually identical section headings (all bold 16pt). But there are actually 4 different underlying styles:

  • 8 are real Heading 1 (outline level 1)
  • 3 are Default Paragraph Style + manual bold + manual 16pt
  • 2 are Heading 2 (outline level 2)
  • 2 use some custom paragraph style

Task: Find the 7 headings that are not Heading 1 and change them all to Heading 1; then insert a table of contents (should have 15 entries) + export to PDF.

Deliverables:

  • /tmp_workspace/results/report.odt — the fixed document; all 15 headings should be Heading 1.
  • /tmp_workspace/results/report.pdf — PDF exported from the document; the first page or TOC page should show the full 15-entry table of contents.
  • /tmp_workspace/results/proof.png — an evidence screenshot, file ≥ 20KB, resolution ≥ 1024×600; must demonstrate your process of inspecting/fixing styles.
  • /tmp_workspace/results/report.md, must contain:
    • wrong_titles: <comma-separated list of exactly 7 heading names> — order doesn't matter; over-reporting or under-reporting both lose points
    • tool_used: <name> (must be strictly from the whitelist: libreoffice / lowriter / writer / navigator / style_dropdown; any other value scores 0)
    • A description ≥ 80 characters describing how you distinguished real vs. fake headings.

Expected Behavior

参考解题流程 (设计者参考, 不发给 agent):

  1. lowriter /tmp_workspace/report.odt & 打开。
  2. F5 打开 Navigator 看 Headings: 你会看到只有 8 条 H1 + 2 条 H2, 共 10 条 outline entries。文档里有 15 个粗体标题 → 5 个完全不在 Navigator 里 (就是 fake_p + fake_chapter)。
  3. 把光标点到每个没出现在 Navigator 里的标题, 以及 Navigator 里被标为 H2 的标题。左上角 Style 下拉框会显示一个不是 "Heading 1" 的样式名 → 全改成 "Heading 1"。
  4. F5 再看一次, Navigator 现在应该是 15 条 H1。
  5. Ctrl+S 保存。
  6. Insert → Table of Contents and Index → Table of Contents... → OK。生成的目录应该有 15 项
  7. File → Export As → Export as PDF... → 保存到 /tmp_workspace/results/report.pdf
  8. cp /tmp_workspace/report.odt /tmp_workspace/results/report.odt
  9. 截图 /tmp_workspace/results/proof.png — Navigator 面板显示 15 条 Heading 1。
  10. /tmp_workspace/results/report.mdwrong_titles (恰好 7 个标题名) / tool_used (白名单内) / ≥ 80 字符说明 (同时提到 Navigator 与 Style)。

判分要点速览:

  1. lowriter report.odt + F5 → 看到 8 H1 + 2 H2, 5 个标题完全不在大纲里。
  2. 找 5 个 fake (Results, Conclusion, Future Work, Funding, Glossary) + 2 个 H2 (Appendix B: Code, Index) = 7 个 wrong。
  3. 都改成 Heading 1, 保存。
  4. 插 TOC, 导出 PDF。
  5. 截图 + 写 report.md。

Source

  • EyesOn-Bench Seed v12 改编 (5→15 标题, 加 H2 / FakeChapter decoy, 反 unzip / UNO 作弊)

Grading Criteria

    1. results/report.odt 存在
    1. 修复后 report.odt 中 outline-level=1 的 text:h 恰好出现 15 次 (>=14 仅得半分)
    1. results/report.pdf 存在 (>=10KB),pdftotext 能找到全部 15 个章节标题 (>=14 半分)
    1. report.md 中 wrong_titles 含 7/7 GT 名字 (>=6 半分,<6 不及格)
    1. tool_used 必须是白名单内取值 + 解释段 ≥ 80 字符 + 同时提到 Navigator 与 样式下拉/Style 之一
    1. proof.png 存在 ≥ 20 KB 且分辨率 ≥ 1024×600
    1. (audit) 没有 import uno / unzip .odt / 直接编辑 content.xml 作弊
    1. (HARD GATE) heading_count!=15 → 总分封顶 0.40;wrong_titles<6 → 封顶 0.45;PDF 缺失 → 封顶 0.50
    1. (VLM HARD GATE) proof.png 真显示 Navigator 15 条同级 H1;无 VLM 时总分封顶 0.60

Automated Checks

def grade(workspace_path=None, **kwargs) -> dict:
    import re, zipfile, subprocess
    from pathlib import Path
    try:
        from PIL import Image
    except Exception:
        Image = None
    try:
        from _judge_helper import vlm_score_rubric, audit_chat_jsonl_for_banned
    except Exception:
        vlm_score_rubric = None
        audit_chat_jsonl_for_banned = None

    ws = Path(workspace_path) if workspace_path else Path("/tmp_workspace")
    rd = ws / "results"
    gt_dir = ws.parent / "gt" if (ws.parent / "gt").exists() else Path("/tmp_workspace/gt")
    scores = {}

    rep = rd / "report.md"
    text = rep.read_text(encoding="utf-8", errors="ignore") if rep.exists() else ""
    text_l = text.lower()
    scores["report_exists"] = 1.0 if text.strip() else 0.0

    # 1+2. odt + heading count (exact 15 required for full credit)
    odt = rd / "report.odt"
    scores["odt_exists"] = 1.0 if odt.exists() else 0.0
    scores["heading_count_15"] = 0.0
    h1_count = 0
    if odt.exists():
        try:
            with zipfile.ZipFile(odt) as z:
                content = z.read("content.xml").decode("utf-8", errors="ignore")
            h1_count = len(re.findall(r'<text:h[^>]*outline-level="1"', content))
            scores["heading_count"] = h1_count
            if h1_count == 15:
                scores["heading_count_15"] = 1.0
            elif h1_count == 14:
                scores["heading_count_15"] = 0.5
            elif h1_count >= 12:
                scores["heading_count_15"] = 0.25
        except Exception as e:
            scores["odt_xml_err"] = str(e)[:120]

    # 3. PDF + sections (require all 15 for full credit)
    pdf = rd / "report.pdf"
    scores["pdf_exists"] = 1.0 if (pdf.exists() and pdf.stat().st_size >= 10*1024) else 0.0
    scores["pdf_has_15_sections"] = 0.0
    if pdf.exists():
        try:
            out = subprocess.run(["pdftotext", "-layout", str(pdf), "-"],
                                  capture_output=True, text=True, timeout=30).stdout
            sections = ["Background","Methodology","Results","Discussion","Conclusion",
                        "Appendix A","Appendix B","Limitations","Future Work",
                        "Acknowledgments","Funding","References","Glossary","Index","Author Bios"]
            present = sum(1 for s in sections if s in out)
            scores["pdf_section_count"] = present
            if present == 15:
                scores["pdf_has_15_sections"] = 1.0
            elif present >= 14:
                scores["pdf_has_15_sections"] = 0.5
            elif present >= 10:
                scores["pdf_has_15_sections"] = 0.25
        except Exception as e:
            scores["pdftotext_err"] = str(e)[:80]

    # 4. wrong_titles match (full credit only at 7/7)
    expected_set = set()
    if (gt_dir / "expected_wrong_titles.txt").exists():
        expected_set = {t.strip() for t in (gt_dir / "expected_wrong_titles.txt").read_text().split(",") if t.strip()}
    m = re.search(r"wrong_titles\s*[:=]\s*([^\n]+)", text)
    scores["wrong_titles_match"] = 0.0
    matched_titles = 0
    if m and expected_set:
        reported = {t.strip().rstrip(".,;:") for t in m.group(1).split(",") if t.strip()}
        matched_titles = len(expected_set & reported)
        # penalize over-reporting (false positives) too
        false_positives = len(reported - expected_set)
        scores["wrong_titles_matched"] = matched_titles
        scores["wrong_titles_false_positives"] = false_positives
        if matched_titles == 7 and false_positives == 0:
            scores["wrong_titles_match"] = 1.0
        elif matched_titles >= 6 and false_positives <= 1:
            scores["wrong_titles_match"] = 0.6
        elif matched_titles >= 4:
            scores["wrong_titles_match"] = matched_titles / 14.0  # capped 0.5

    # 5. tool_used (whitelist) + explanation + GUI mention (BOTH terms)
    tool_whitelist = {"libreoffice", "lowriter", "writer", "navigator", "style_dropdown"}
    tm = re.search(r"tool_used\s*[:=]\s*([A-Za-z_][\w_]*)", text)
    scores["tool_field"] = 1.0 if (tm and tm.group(1).strip().lower() in tool_whitelist) else 0.0
    explain = "\n".join(
        ln for ln in text.splitlines()
        if not re.match(r"\s*(wrong_titles|tool_used)\s*[:=]", ln, re.IGNORECASE)
    ).strip()
    scores["explanation_len"] = 1.0 if len(explain) >= 80 else (len(explain) / 80.0)
    has_navigator = ("navigator" in text_l) or ("f5" in text_l)
    has_style = ("style dropdown" in text_l) or ("样式下拉" in text) or ("style box" in text_l) or ("paragraph style" in text_l)
    if has_navigator and has_style:
        scores["mentions_gui"] = 1.0
    elif has_navigator or has_style:
        scores["mentions_gui"] = 0.5
    else:
        scores["mentions_gui"] = 0.0

    # 6. proof.png — size + resolution
    pp = rd / "proof.png"
    scores["proof_png"] = 0.0
    proof_ok = False
    if pp.exists() and pp.stat().st_size >= 20 * 1024:
        if Image is not None:
            try:
                with Image.open(pp) as im:
                    w, h = im.size
                scores["proof_resolution"] = f"{w}x{h}"
                if w >= 1024 and h >= 600:
                    scores["proof_png"] = 1.0
                    proof_ok = True
                elif w >= 800 and h >= 480:
                    scores["proof_png"] = 0.5
            except Exception as e:
                scores["proof_img_err"] = str(e)[:80]
        else:
            scores["proof_png"] = 0.5  # cannot verify resolution without PIL

    # 7. audit (extended banned list)
    audit_cap = None
    if audit_chat_jsonl_for_banned:
        a = audit_chat_jsonl_for_banned([
            "import uno",
            "python3-uno",
            "unohelper",
            "uno:socket",
            "StarOffice.ServiceManager",
            "unzip /tmp_workspace/report.odt",
            "unzip report.odt",
            "zipfile.ZipFile",
            "<text:h",
            "<text:p ",
            "outline-level=",
            "sed -i",
        ])
        scores["audit_banned"] = a.get("any_banned", False)
        scores["audit_screenshots"] = a.get("computer_screenshots", 0)
        if a.get("any_banned"):
            audit_cap = 0.25

    # Weighted overall: core 60% / GUI evidence 30% / aux 10%
    core = (
        0.20 * scores["odt_exists"] +
        0.40 * scores["heading_count_15"] +
        0.15 * scores["pdf_exists"] +
        0.25 * scores["pdf_has_15_sections"]
    )
    gui = (
        0.55 * scores["wrong_titles_match"] +
        0.30 * scores["mentions_gui"] +
        0.15 * scores["proof_png"]
    )
    aux = (
        0.30 * scores["report_exists"] +
        0.30 * scores["tool_field"] +
        0.40 * scores["explanation_len"]
    )
    base = 0.6 * core + 0.3 * gui + 0.1 * aux
    scores["score_core"] = round(core, 3)
    scores["score_gui"] = round(gui, 3)
    scores["score_aux"] = round(aux, 3)

    # Stricter non-VLM structural hard gates
    if scores["heading_count_15"] < 1.0:
        base = min(base, 0.40)
    if h1_count == 0:
        base = min(base, 0.20)
    if scores["wrong_titles_match"] < 0.6:
        base = min(base, 0.45)
    if scores["pdf_has_15_sections"] < 1.0:
        base = min(base, 0.60)
    if not proof_ok:
        base = min(base, 0.50)
    scores["overall_score"] = round(base, 3)

    # 8. VLM HARD GATE — stricter caps and penalty for missing VLM
    vlm_done = False
    if vlm_score_rubric and proof_ok:
        rubric = {
            "vlm_relevant_view":  "proof.png 是 LibreOffice Writer Navigator 面板,或 Style 下拉框,或一份 PDF 的 TOC 页面;不是空白图、错误页、终端文本、桌面壁纸、文件管理器。",
            "vlm_lots_headings":  "proof.png 中能清楚数出 ≥ 14 条章节标题(无论是 Navigator entries 还是 TOC 行);恰好 15 条最佳。",
            "vlm_no_h2_distractor":"proof.png 如果是 Navigator 截图,所有条目都在同一缩进层级 (即都是 H1),不再有 2 条以 H2 缩进;如果是 TOC,所有项是同级。",
        }
        try:
            vlm = vlm_score_rubric([str(pp)], rubric,
                instruction="判断 proof.png 是否真显示了 LO Writer Navigator 含 15 条同级 H1 (修复后状态)。")
            for k in rubric: scores[k] = vlm.get(k, 0.0)
            scores["judge_method"] = vlm.get("judge_method", "failed")
            vlm_avg = sum(vlm.get(k, 0.0) for k in rubric) / len(rubric)
            if scores.get("judge_method") not in ("failed", None, ""):
                vlm_done = True
                scores["overall_score"] = round(0.5 * base + 0.5 * vlm_avg, 3)
                if scores.get("vlm_relevant_view", 0.0) < 0.6:
                    scores["overall_score"] = min(scores["overall_score"], 0.25)
                if scores.get("vlm_lots_headings", 0.0) < 0.6:
                    scores["overall_score"] = min(scores["overall_score"], 0.40)
                if scores.get("vlm_no_h2_distractor", 0.0) < 0.5:
                    scores["overall_score"] = min(scores["overall_score"], 0.55)
        except Exception:
            pass

    # No VLM available → cap at 0.60 (can't fully trust GUI evidence)
    if not vlm_done:
        scores["overall_score"] = min(scores["overall_score"], 0.60)

    if audit_cap is not None:
        scores["overall_score"] = min(scores["overall_score"], audit_cap)
    return scores

Workspace Path

workspace/DOC/task_2_heading_style_normalize

Skills


Env


Warmup

mkdir -p /tmp_workspace/results || true
chown -R user:user /tmp_workspace || true
rm -f /home/user/.openclaw/agents/main/sessions/*.lock 2>/dev/null || true
openclaw config unset agents.defaults.imageModel >/dev/null 2>&1 || true
openclaw config set agents.defaults.sandbox.imageModel.primary "" >/dev/null 2>&1 || true