feat: Module C v5/v6 training complete, ablations, SOTA baselines, paper updates

- Module C: BC+PPO training v5/v6 done; eval results in experiments/eval_intervention_v{5,6}.json - Reward: v5 label-aligned constrained reward (code/src/rl/reward.py) - Ablations: Module B (history_r, response_only, full) + Module C (wo_category_reward) - SOTA baselines: WildGuard and ShieldGemma2b eval scripts and results - Paper: update sections 05–08 (Module B/C description, experiments table, discussion) - Docs: add record.md (change log), update state.md and exp.md; retire change.md - Tools: add html-to-ppt utilities and run_shieldgemma2b.sh - Configs: add ablation YAML configs for Module B and C - Cleanup: remove stale reference/ PNG screenshots Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-05-20 14:24:09 +08:00
parent 6d61a950f1
commit 52ba43f08d
55 changed files with 8239 additions and 1244 deletions
--- a/tools/html_to_ppt_capture.js
+++ b/tools/html_to_ppt_capture.js
@@ -0,0 +1,114 @@
+const fs = require('fs/promises');
+const path = require('path');
+const { chromium } = require('playwright');
+
+const input = process.argv[2];
+const outputDir = process.argv[3];
+const presenterNames = ['Zhihao Zhao', 'Zipeng Wang', 'Jiuqi Feng', 'Siyuan Zhang'];
+
+function baseWithoutHash(url) {
+  const hashIndex = url.indexOf('#');
+  return hashIndex >= 0 ? url.slice(0, hashIndex) : url;
+}
+
+async function applyRequestedEdits(page) {
+  await page.evaluate((names) => {
+    const visible = Array.from(document.querySelectorAll('section')).find((section) => {
+      const style = getComputedStyle(section);
+      return style.visibility !== 'hidden' && style.opacity !== '0';
+    });
+    if (!visible) return;
+
+    if ((visible.getAttribute('data-label') || '').startsWith('01')) {
+      for (const el of Array.from(visible.querySelectorAll('div'))) {
+        const text = (el.innerText || '').replace(/\s+/g, ' ').trim();
+        if (text === 'PRESENTED [ Presentation Date ] 4-person team walkthrough · ~10 min') {
+          el.style.display = 'none';
+        }
+      }
+    }
+
+    if ((visible.getAttribute('data-label') || '').startsWith('15')) {
+      const placeholders = Array.from(visible.querySelectorAll('div'))
+        .filter((el) => (el.innerText || '').trim() === '[ name ]');
+      placeholders.forEach((el, index) => {
+        if (names[index]) {
+          el.textContent = names[index];
+        }
+      });
+    }
+  }, presenterNames);
+}
+
+(async () => {
+  if (!input || !outputDir) {
+    throw new Error('Usage: node html_to_ppt_capture.js <file-url> <output-dir>');
+  }
+
+  await fs.mkdir(outputDir, { recursive: true });
+
+  const launchOptions = { headless: true };
+  if (process.env.BROWSER_EXE) {
+    launchOptions.executablePath = process.env.BROWSER_EXE;
+  }
+
+  const browser = await chromium.launch(launchOptions);
+  const firstPage = await browser.newPage({
+    viewport: { width: 1920, height: 1080 },
+    deviceScaleFactor: 1,
+  });
+
+  const baseUrl = baseWithoutHash(input);
+  await firstPage.goto(`${baseUrl}#1`, { waitUntil: 'load', timeout: 30000 });
+  await firstPage.waitForSelector('section', { state: 'attached', timeout: 10000 });
+  await firstPage.evaluate(() => document.fonts && document.fonts.ready);
+  await firstPage.waitForTimeout(500);
+
+  const slideCount = await firstPage.evaluate(() => document.querySelectorAll('section').length);
+  await firstPage.close();
+  const labels = [];
+
+  for (let i = 1; i <= slideCount; i += 1) {
+    const page = await browser.newPage({
+      viewport: { width: 1920, height: 1080 },
+      deviceScaleFactor: 1,
+    });
+    await page.goto(`${baseUrl}#${i}`, { waitUntil: 'load', timeout: 30000 });
+    await page.waitForSelector('section', { state: 'attached', timeout: 10000 });
+    await page.evaluate(() => document.fonts && document.fonts.ready);
+    await page.waitForTimeout(900);
+    await applyRequestedEdits(page);
+    await page.waitForTimeout(100);
+
+    const { label, clip } = await page.evaluate(() => {
+      const visible = Array.from(document.querySelectorAll('section')).find((section) => {
+        const style = getComputedStyle(section);
+        return style.visibility !== 'hidden' && style.opacity !== '0';
+      });
+      const rect = visible?.getBoundingClientRect();
+      return {
+        label: visible?.getAttribute('data-label') || document.body.innerText.split('\n')[0] || '',
+        clip: rect
+          ? {
+              x: Math.max(0, Math.round(rect.x)),
+              y: Math.max(0, Math.round(rect.y)),
+              width: Math.round(rect.width),
+              height: Math.round(rect.height),
+            }
+          : null,
+      };
+    });
+    labels.push({ index: i, label });
+
+    const file = path.join(outputDir, `slide_${String(i).padStart(2, '0')}.png`);
+    await page.screenshot({ path: file, fullPage: false, clip: clip || undefined });
+    console.log(`captured ${i}/${slideCount}: ${label}`);
+    await page.close();
+  }
+
+  await fs.writeFile(path.join(outputDir, 'slides.json'), JSON.stringify(labels, null, 2), 'utf8');
+  await browser.close();
+})().catch((err) => {
+  console.error(err);
+  process.exit(1);
+});
--- a/tools/html_to_ppt_check_placeholders.js
+++ b/tools/html_to_ppt_check_placeholders.js
@@ -0,0 +1,89 @@
+const { chromium } = require('playwright');
+
+const input = process.argv[2];
+const presenterNames = ['Zhihao Zhao', 'Zipeng Wang', 'Jiuqi Feng', 'Siyuan Zhang'];
+
+function baseWithoutHash(url) {
+  const hashIndex = url.indexOf('#');
+  return hashIndex >= 0 ? url.slice(0, hashIndex) : url;
+}
+
+async function applyRequestedEdits(page) {
+  await page.evaluate((names) => {
+    const visible = Array.from(document.querySelectorAll('section')).find((section) => {
+      const style = getComputedStyle(section);
+      return style.visibility !== 'hidden' && style.opacity !== '0';
+    });
+    if (!visible) return;
+
+    if ((visible.getAttribute('data-label') || '').startsWith('01')) {
+      for (const el of Array.from(visible.querySelectorAll('div'))) {
+        const text = (el.innerText || '').replace(/\s+/g, ' ').trim();
+        if (text === 'PRESENTED [ Presentation Date ] 4-person team walkthrough · ~10 min') {
+          el.style.display = 'none';
+        }
+      }
+    }
+
+    if ((visible.getAttribute('data-label') || '').startsWith('15')) {
+      const placeholders = Array.from(visible.querySelectorAll('div'))
+        .filter((el) => (el.innerText || '').trim() === '[ name ]');
+      placeholders.forEach((el, index) => {
+        if (names[index]) {
+          el.textContent = names[index];
+        }
+      });
+    }
+  }, presenterNames);
+}
+
+(async () => {
+  if (!input) {
+    throw new Error('Usage: node html_to_ppt_check_placeholders.js <file-url>');
+  }
+
+  const launchOptions = { headless: true };
+  if (process.env.BROWSER_EXE) {
+    launchOptions.executablePath = process.env.BROWSER_EXE;
+  }
+
+  const browser = await chromium.launch(launchOptions);
+  const baseUrl = baseWithoutHash(input);
+  const findings = [];
+  const nameChecks = [];
+
+  for (let i = 1; i <= 15; i += 1) {
+    const page = await browser.newPage({ viewport: { width: 1920, height: 1080 } });
+    await page.goto(`${baseUrl}#${i}`, { waitUntil: 'load', timeout: 30000 });
+    await page.waitForSelector('section', { state: 'attached', timeout: 10000 });
+    await page.waitForTimeout(500);
+    await applyRequestedEdits(page);
+    const result = await page.evaluate((names) => {
+      const visible = Array.from(document.querySelectorAll('section')).find((section) => {
+        const style = getComputedStyle(section);
+        return style.visibility !== 'hidden' && style.opacity !== '0';
+      });
+      const label = visible?.getAttribute('data-label') || '';
+      const text = visible?.innerText || '';
+      const placeholderMatches = text.match(/\[[^\]]+\]|Presentation Date|Slide Number|placeholder|PRESENTED\s+\[/gi) || [];
+      return {
+        label,
+        placeholderMatches,
+        presenterNamesPresent: names.map((name) => text.includes(name)),
+      };
+    }, presenterNames);
+    if (result.placeholderMatches.length) {
+      findings.push({ slide: i, label: result.label, matches: result.placeholderMatches });
+    }
+    if (i === 15) {
+      nameChecks.push(...result.presenterNamesPresent);
+    }
+    await page.close();
+  }
+
+  await browser.close();
+  console.log(JSON.stringify({ findings, slide15NamesPresent: nameChecks }, null, 2));
+})().catch((err) => {
+  console.error(err);
+  process.exit(1);
+});
--- a/tools/html_to_ppt_make.py
+++ b/tools/html_to_ppt_make.py
@@ -0,0 +1,68 @@
+import json
+import sys
+from pathlib import Path
+
+from PIL import Image, ImageOps
+from pptx import Presentation
+from pptx.util import Inches
+
+
+def main() -> None:
+    if len(sys.argv) != 3:
+        raise SystemExit("Usage: python html_to_ppt_make.py <image-dir> <output-pptx>")
+
+    image_dir = Path(sys.argv[1])
+    output_pptx = Path(sys.argv[2])
+    output_pptx.parent.mkdir(parents=True, exist_ok=True)
+
+    labels_path = image_dir / "slides.json"
+    labels = {}
+    if labels_path.exists():
+        labels = {item["index"]: item["label"] for item in json.loads(labels_path.read_text(encoding="utf-8"))}
+
+    image_paths = sorted(image_dir.glob("slide_*.png"))
+    if not image_paths:
+        raise SystemExit(f"No slide PNGs found in {image_dir}")
+
+    prs = Presentation()
+    prs.slide_width = Inches(13.333333)
+    prs.slide_height = Inches(7.5)
+
+    blank_layout = prs.slide_layouts[6]
+    for image_path in image_paths:
+        slide_num = int(image_path.stem.split("_")[-1])
+        slide = prs.slides.add_slide(blank_layout)
+        picture = slide.shapes.add_picture(
+            str(image_path),
+            0,
+            0,
+            width=prs.slide_width,
+            height=prs.slide_height,
+        )
+        picture.name = labels.get(slide_num, image_path.stem)
+
+    prs.save(output_pptx)
+
+    montage_path = output_pptx.with_suffix(".preview.png")
+    thumbs = []
+    for image_path in image_paths:
+        img = Image.open(image_path).convert("RGB")
+        img.thumbnail((384, 216), Image.Resampling.LANCZOS)
+        thumbs.append(ImageOps.expand(img, border=2, fill=(30, 30, 30)))
+
+    cols = 5
+    rows = (len(thumbs) + cols - 1) // cols
+    montage = Image.new("RGB", (cols * 388, rows * 220), (12, 15, 24))
+    for idx, thumb in enumerate(thumbs):
+        x = (idx % cols) * 388
+        y = (idx // cols) * 220
+        montage.paste(thumb, (x, y))
+    montage.save(montage_path)
+
+    print(f"pptx={output_pptx}")
+    print(f"preview={montage_path}")
+    print(f"slides={len(image_paths)}")
+
+
+if __name__ == "__main__":
+    main()
--- a/tools/html_to_ppt_probe.js
+++ b/tools/html_to_ppt_probe.js
@@ -0,0 +1,43 @@
+const { chromium } = require('playwright');
+
+const input = process.argv[2];
+
+(async () => {
+  const launchOptions = { headless: true };
+  if (process.env.BROWSER_EXE) {
+    launchOptions.executablePath = process.env.BROWSER_EXE;
+  }
+  const browser = await chromium.launch(launchOptions);
+  const page = await browser.newPage({ viewport: { width: 1920, height: 1080 }, deviceScaleFactor: 1 });
+  await page.goto(input, { waitUntil: 'load', timeout: 30000 });
+  await page.waitForTimeout(1500);
+  await page.waitForSelector('section', { state: 'attached', timeout: 10000 });
+  const info = await page.evaluate(() => {
+    const sections = Array.from(document.querySelectorAll('section'));
+    return {
+      title: document.title,
+      url: location.href,
+      bodyText: document.body.innerText.slice(0, 500),
+      count: sections.length,
+      viewport: { w: innerWidth, h: innerHeight },
+      scroll: { w: document.documentElement.scrollWidth, h: document.documentElement.scrollHeight },
+      sections: sections.map((s, i) => {
+        const r = s.getBoundingClientRect();
+        const cs = getComputedStyle(s);
+        return {
+          index: i + 1,
+          label: s.getAttribute('data-label') || '',
+          rect: { x: r.x, y: r.y, w: r.width, h: r.height },
+          display: cs.display,
+          visibility: cs.visibility,
+          opacity: cs.opacity,
+        };
+      }),
+    };
+  });
+  console.log(JSON.stringify(info, null, 2));
+  await browser.close();
+})().catch((err) => {
+  console.error(err);
+  process.exit(1);
+});
--- a/tools/html_to_ppt_with_notes.js
+++ b/tools/html_to_ppt_with_notes.js
@@ -0,0 +1,75 @@
+const fs = require('fs/promises');
+const path = require('path');
+const { chromium } = require('playwright');
+const pptxgen = require('pptxgenjs');
+
+const inputUrl = process.argv[2];
+const imageDir = process.argv[3];
+const outputPptx = process.argv[4];
+
+async function extractNotes(url) {
+  const launchOptions = { headless: true };
+  if (process.env.BROWSER_EXE) {
+    launchOptions.executablePath = process.env.BROWSER_EXE;
+  }
+
+  const browser = await chromium.launch(launchOptions);
+  const page = await browser.newPage({ viewport: { width: 1920, height: 1080 } });
+  const baseUrl = url.includes('#') ? url.slice(0, url.indexOf('#')) : url;
+  await page.goto(`${baseUrl}#1`, { waitUntil: 'load', timeout: 30000 });
+  await page.waitForSelector('#speaker-notes', { state: 'attached', timeout: 10000 });
+  const notes = await page.evaluate(() => JSON.parse(document.getElementById('speaker-notes').textContent));
+  await browser.close();
+  return notes;
+}
+
+(async () => {
+  if (!inputUrl || !imageDir || !outputPptx) {
+    throw new Error('Usage: node html_to_ppt_with_notes.js <file-url> <image-dir> <output-pptx>');
+  }
+
+  const notes = await extractNotes(inputUrl);
+  const images = (await fs.readdir(imageDir))
+    .filter((name) => /^slide_\d+\.png$/i.test(name))
+    .sort()
+    .map((name) => path.join(imageDir, name));
+
+  if (notes.length !== images.length) {
+    throw new Error(`Notes count (${notes.length}) does not match image count (${images.length})`);
+  }
+
+  await fs.mkdir(path.dirname(outputPptx), { recursive: true });
+  await fs.writeFile(
+    path.join(path.dirname(outputPptx), 'Generative_Image_Dynamics.notes.json'),
+    JSON.stringify(notes.map((text, index) => ({ slide: index + 1, text })), null, 2),
+    'utf8',
+  );
+
+  const pptx = new pptxgen();
+  pptx.layout = 'LAYOUT_WIDE';
+  pptx.author = 'Codex';
+  pptx.subject = 'Generative Image Dynamics — CVPR 2024';
+  pptx.title = 'Generative Image Dynamics';
+  pptx.company = '';
+  pptx.lang = 'en-US';
+  pptx.theme = {
+    headFontFace: 'Aptos',
+    bodyFontFace: 'Aptos',
+    lang: 'en-US',
+  };
+
+  for (let i = 0; i < images.length; i += 1) {
+    const slide = pptx.addSlide();
+    slide.background = { color: '000000' };
+    slide.addImage({ path: images[i], x: 0, y: 0, w: 13.333333, h: 7.5 });
+    slide.addNotes(notes[i]);
+  }
+
+  await pptx.writeFile({ fileName: outputPptx });
+  console.log(`pptx=${outputPptx}`);
+  console.log(`slides=${images.length}`);
+  console.log(`notes=${notes.length}`);
+})().catch((err) => {
+  console.error(err);
+  process.exit(1);
+});
--- a/tools/run_shieldgemma2b.sh
+++ b/tools/run_shieldgemma2b.sh
@@ -0,0 +1,45 @@
+#!/bin/bash
+# 登录 HF → 下载 ShieldGemma-2B → 运行评估，全程写入日志
+set -e
+
+PROJ=/root/siton-data-2849d4ce327c4ccfb233ce33868fe7fe/zsy/CompanionGuard-RL
+MODEL_DIR=/root/siton-data-2849d4ce327c4ccfb233ce33868fe7fe/zsy/shieldgemma-2b
+PY=/opt/conda/envs/dlapo-py310-cu128/bin/python
+HF=/opt/conda/envs/dlapo-py310-cu128/bin/hf
+LOG=$PROJ/experiments/run_shieldgemma2b_$(date +%Y%m%d_%H%M%S).log
+
+mkdir -p $PROJ/experiments
+
+# 从这里开始把所有输出重定向到日志文件
+exec > "$LOG" 2>&1
+
+echo "=== $(date) START ==="
+echo "PROJ=$PROJ"
+echo "MODEL_DIR=$MODEL_DIR"
+
+# 代理（服务器无外网）
+export http_proxy=http://127.0.0.1:7890
+export https_proxy=http://127.0.0.1:7890
+
+echo ""
+echo "--- [1/3] HuggingFace Login ---"
+$HF auth login --token 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
+
+echo ""
+echo "--- [2/3] Downloading google/shieldgemma-2b ---"
+$HF download google/shieldgemma-2b \
+  --local-dir "$MODEL_DIR"
+
+echo ""
+echo "--- [3/3] Running evaluation on CompanionRisk-Bench test set ---"
+cd "$PROJ"
+export PYTHONPATH="$PROJ"
+CUDA_VISIBLE_DEVICES=0 $PY scripts/eval_sota_baselines.py \
+  --model shieldgemma2b \
+  --model-path "$MODEL_DIR" \
+  --test-data data/processed/CompanionRisk-Bench/test.jsonl \
+  --output experiments/eval_sota_shieldgemma2b.json
+
+echo ""
+echo "=== $(date) DONE ==="
+echo "Result: $PROJ/experiments/eval_sota_shieldgemma2b.json"
--- a/tools/scrub_pptx_placeholders.py
+++ b/tools/scrub_pptx_placeholders.py
@@ -0,0 +1,44 @@
+import re
+import sys
+import zipfile
+from pathlib import Path
+
+
+PLACEHOLDER_PATTERNS = [
+    r"<p:sp\b(?:(?!</p:sp>).)*?<p:ph\b[^>]*(?:type=\"sldNum\"|type=\"dt\"|type=\"ftr\")[^>]*/>(?:(?!</p:sp>).)*?</p:sp>",
+    r"<p:sp\b(?:(?!</p:sp>).)*?<a:t>Slide Number</a:t>(?:(?!</p:sp>).)*?</p:sp>",
+    r"<p:sp\b(?:(?!</p:sp>).)*?<a:t>Date Placeholder</a:t>(?:(?!</p:sp>).)*?</p:sp>",
+    r"<p:sp\b(?:(?!</p:sp>).)*?<a:t>Footer</a:t>(?:(?!</p:sp>).)*?</p:sp>",
+]
+
+
+def scrub_xml(xml: str) -> str:
+    for pattern in PLACEHOLDER_PATTERNS:
+        xml = re.sub(pattern, "", xml, flags=re.DOTALL)
+    xml = re.sub(r"\s+sldNum=\"[^\"]*\"", "", xml)
+    return xml
+
+
+def main() -> None:
+    if len(sys.argv) != 3:
+        raise SystemExit("Usage: python scrub_pptx_placeholders.py <input.pptx> <output.pptx>")
+
+    src = Path(sys.argv[1])
+    dst = Path(sys.argv[2])
+    dst.parent.mkdir(parents=True, exist_ok=True)
+
+    with zipfile.ZipFile(src, "r") as zin, zipfile.ZipFile(dst, "w", zipfile.ZIP_DEFLATED) as zout:
+        for item in zin.infolist():
+            data = zin.read(item.filename)
+            if item.filename.endswith(".xml") and (
+                item.filename.startswith("ppt/notesSlides/")
+                or item.filename.startswith("ppt/notesMasters/")
+                or item.filename.startswith("ppt/slideMasters/")
+            ):
+                text = data.decode("utf-8")
+                data = scrub_xml(text).encode("utf-8")
+            zout.writestr(item, data)
+
+
+if __name__ == "__main__":
+    main()