| <!doctype html> |
| <html lang="ko"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1"><title>TAC-GUI Trajectory Branch v1</title><style> |
| :root{--bg:#f7f8fb;--ink:#111827;--muted:#5b6472;--line:#d9dee8;--panel:#fff;--gold:#b7791f;--bad:#64748b;--accent:#1d4ed8;--ok:#15803d} |
| *{box-sizing:border-box}body{margin:0;background:var(--bg);color:var(--ink);font-family:-apple-system,BlinkMacSystemFont,"Segoe UI",Pretendard,"Noto Sans KR",sans-serif;line-height:1.5}header{padding:34px 42px 24px;background:#fff;border-bottom:1px solid var(--line)}h1{margin:0 0 8px;font-size:30px;letter-spacing:-.02em}.lead{max-width:980px;color:var(--muted);margin:0}.wrap{padding:28px 42px 70px}.stats{display:grid;grid-template-columns:repeat(4,minmax(0,1fr));gap:12px;margin:22px 0}.stat{background:#fff;border:1px solid var(--line);border-radius:8px;padding:14px}.stat strong{display:block;font-size:24px}.stat span{font-size:13px;color:var(--muted)}.section{margin-top:34px}.step{background:#fff;border:1px solid var(--line);border-radius:8px;padding:20px;margin:20px 0}.step-head{display:flex;justify-content:space-between;gap:20px;align-items:start;margin-bottom:16px}.eyebrow{text-transform:uppercase;letter-spacing:.12em;font-size:11px;color:var(--accent);font-weight:700}.intent{font-size:18px;font-weight:700;margin:4px 0}.path{font-family:ui-monospace,SFMono-Regular,Menlo,monospace;color:var(--muted);font-size:12px}.grid{display:grid;grid-template-columns:360px 1fr;gap:18px;align-items:start}.current img,.card img{width:100%;border:1px solid var(--line);border-radius:6px;background:#111}.current h3,.branches h3{margin:0 0 10px;font-size:14px;color:var(--muted)}.branch-grid{display:grid;grid-template-columns:repeat(2,minmax(0,1fr));gap:14px}.card{border:1px solid var(--line);border-radius:8px;padding:12px;background:#fbfcff}.card.gold{border-color:#f0c36a;background:#fffaf0}.badge{display:inline-flex;border-radius:999px;padding:2px 8px;font-size:11px;font-weight:700;background:#e5e7eb;color:#374151}.badge.gold{background:#fef3c7;color:#92400e}.badge.valid{background:#dcfce7;color:#166534}.cmd{font-family:ui-monospace,SFMono-Regular,Menlo,monospace;background:#f1f5f9;padding:8px;border-radius:6px;white-space:pre-wrap;font-size:12px;margin:8px 0}.meta{font-size:12px;color:var(--muted);display:flex;gap:8px;flex-wrap:wrap;margin:6px 0 10px}.json{white-space:pre-wrap;background:#0f172a;color:#dbeafe;padding:12px;border-radius:6px;font-size:11px;max-height:260px;overflow:auto}.note{background:#eff6ff;border:1px solid #bfdbfe;border-radius:8px;padding:14px;margin-top:14px;color:#1e3a8a}@media(max-width:1100px){.grid{grid-template-columns:1fr}.branch-grid{grid-template-columns:1fr}.stats{grid-template-columns:repeat(2,1fr)}} |
| </style></head> |
| <body> |
| <header> |
| <div class="eyebrow">TAC-GUI trajectory branching dataset v1</div> |
| <h1>AgentNet 스타일: 긴 task, local code, 그리고 같은 state의 branches</h1> |
| <p class="lead">이 샘플은 기존 AgentNet/OpenCUA처럼 task는 long-horizon으로 유지하고, 각 row의 <code>## Code</code>는 다음 local action 하나만 수행하도록 구성했다. 동시에 각 step마다 같은 current state에서 gold/negative action branches를 제공한다. 모든 branch는 current와 next screenshot의 pixel diff가 0보다 큰 경우만 남겼다.</p> |
| </header> |
| <main class="wrap"> |
| <div class="stats"> |
| <div class="stat"><strong>4</strong><span>gold trajectory steps</span></div> |
| <div class="stat"><strong>16</strong><span>valid branch transitions</span></div> |
| <div class="stat"><strong>0</strong><span>dropped no-op branches</span></div> |
| <div class="stat"><strong>4</strong><span>branches per state</span></div> |
| </div> |
| <div class="note"><strong>Task:</strong> In the Linux terminal, inspect docs/guide.txt, return to the root folder, then enter the data folder.<br>핵심은 “task는 여러 단계지만, 각 row의 code는 지금 state에서 해야 할 다음 한 동작”이라는 점이다. 그래서 WM 학습 row는 <code>s_t + candidate code -> s_t+1</code>이고, planning/policy 평가에서는 이 row들이 trajectory prefix를 따라 이어진다.</div> |
| <section class="section"><h2>Gold trajectory with branches</h2> |
| <section class="step" id="step-0"> |
| <div class="step-head"> |
| <div> |
| <div class="eyebrow">Step 0 · same current state, multiple candidate local actions</div> |
| <div class="intent">Open the docs folder first.</div> |
| <div class="path">base_state=traj0/r · gold=cd docs</div> |
| </div> |
| </div> |
| <div class="grid"> |
| <div class="current"><h3>Current state s_t</h3><img src="assets/step00_traj0_r_current_s000000.png" alt="current step 0"></div> |
| <div class="branches"><h3>Candidate branches: s_t + action -> s_t+1</h3><div class="branch-grid"> |
| <article class="card gold"> |
| <div><span class="badge gold">GOLD</span> <span class="badge valid">valid diff</span></div> |
| <div class="meta"><span>gold</span><span>diff 0.11%</span><span>s00_gold_cd_docs</span></div> |
| <img src="assets/step00_traj0_r_s00_gold_cd_docs_s000107.png" alt="next state s00_gold_cd_docs"> |
| <div class="cmd">pyautogui.write(message='cd docs') |
| pyautogui.press('enter')</div> |
| <details><summary>OpenCUA row JSON</summary><pre class="json">{ |
| "images": [ |
| "/NHNHOME/data/sanghyeok/qwen-cua/data/tac_gui_traj_branch_v1/images/step00_traj0_r_current_s000000.png" |
| ], |
| "wm_next_image": "/NHNHOME/data/sanghyeok/qwen-cua/data/tac_gui_traj_branch_v1/images/step00_traj0_r_s00_gold_cd_docs_s000107.png", |
| "messages": [ |
| { |
| "role": "system", |
| "content": "You are a GUI agent operating a desktop. For each step, given the user task and the current screenshot, respond in this exact format:\n\n## Thought:\n<brief reasoning about progress and the chosen next local action>\n\n## Action:\n<one-sentence description of what to do>\n\n## Code:\n<pyautogui code; coordinates normalized to [0,1] when clicking>" |
| }, |
| { |
| "role": "user", |
| "content": "Task: In the Linux terminal, inspect docs/guide.txt, return to the root folder, then enter the data folder.\n<image>\nWhat is the next action?" |
| }, |
| { |
| "role": "assistant", |
| "content": "## Thought:\nProgress: Open the docs folder first. The next local terminal action should be 'cd docs'.\n\n## Action:\nType the command 'cd docs' and press Enter.\n\n## Code:\npyautogui.write(message='cd docs')\npyautogui.press('enter')" |
| } |
| ], |
| "meta": { |
| "source": "tac_gui_traj_branch_v1", |
| "base_source": "cf_terminal_traj_fhd", |
| "trajectory_id": "tac_terminal_docs_data_v1", |
| "state_key": "tac_terminal_docs_data_v1/step0", |
| "base_state_key": "traj0/r", |
| "step_index": 0, |
| "task": "In the Linux terminal, inspect docs/guide.txt, return to the root folder, then enter the data folder.", |
| "step_intent": "Open the docs folder first.", |
| "branch_id": "s00_gold_cd_docs", |
| "branch_idx": 1, |
| "branch_type": "gold", |
| "is_gold": true, |
| "gold_command": "cd docs", |
| "command": "cd docs", |
| "action_granularity": "policy_step_local_action", |
| "code_norm": "pyautogui.write(message='cd docs')\npyautogui.press('enter')", |
| "current_image_rel": "images/step00_traj0_r_current_s000000.png", |
| "next_image_rel": "images/step00_traj0_r_s00_gold_cd_docs_s000107.png", |
| "pixel_diff_nonzero": 2275, |
| "pixel_diff_ratio": 0.0010971257716049384 |
| } |
| }</pre></details> |
| </article> |
| <article class="card "> |
| <div><span class="badge">NEGATIVE</span> <span class="badge valid">valid diff</span></div> |
| <div class="meta"><span>wrong_navigation</span><span>diff 0.11%</span><span>s00_neg_cd_data</span></div> |
| <img src="assets/step00_traj0_r_s00_neg_cd_data_s000001.png" alt="next state s00_neg_cd_data"> |
| <div class="cmd">pyautogui.write(message='cd data') |
| pyautogui.press('enter')</div> |
| <details><summary>OpenCUA row JSON</summary><pre class="json">{ |
| "images": [ |
| "/NHNHOME/data/sanghyeok/qwen-cua/data/tac_gui_traj_branch_v1/images/step00_traj0_r_current_s000000.png" |
| ], |
| "wm_next_image": "/NHNHOME/data/sanghyeok/qwen-cua/data/tac_gui_traj_branch_v1/images/step00_traj0_r_s00_neg_cd_data_s000001.png", |
| "messages": [ |
| { |
| "role": "system", |
| "content": "You are a GUI agent operating a desktop. For each step, given the user task and the current screenshot, respond in this exact format:\n\n## Thought:\n<brief reasoning about progress and the chosen next local action>\n\n## Action:\n<one-sentence description of what to do>\n\n## Code:\n<pyautogui code; coordinates normalized to [0,1] when clicking>" |
| }, |
| { |
| "role": "user", |
| "content": "Task: In the Linux terminal, inspect docs/guide.txt, return to the root folder, then enter the data folder.\n<image>\nWhat is the next action?" |
| }, |
| { |
| "role": "assistant", |
| "content": "## Thought:\nThis is a plausible but incorrect local action for the same state: 'cd data'.\n\n## Action:\nType the command 'cd data' and press Enter.\n\n## Code:\npyautogui.write(message='cd data')\npyautogui.press('enter')" |
| } |
| ], |
| "meta": { |
| "source": "tac_gui_traj_branch_v1", |
| "base_source": "cf_terminal_traj_fhd", |
| "trajectory_id": "tac_terminal_docs_data_v1", |
| "state_key": "tac_terminal_docs_data_v1/step0", |
| "base_state_key": "traj0/r", |
| "step_index": 0, |
| "task": "In the Linux terminal, inspect docs/guide.txt, return to the root folder, then enter the data folder.", |
| "step_intent": "Open the docs folder first.", |
| "branch_id": "s00_neg_cd_data", |
| "branch_idx": 0, |
| "branch_type": "wrong_navigation", |
| "is_gold": false, |
| "gold_command": "cd docs", |
| "command": "cd data", |
| "action_granularity": "policy_step_local_action", |
| "code_norm": "pyautogui.write(message='cd data')\npyautogui.press('enter')", |
| "current_image_rel": "images/step00_traj0_r_current_s000000.png", |
| "next_image_rel": "images/step00_traj0_r_s00_neg_cd_data_s000001.png", |
| "pixel_diff_nonzero": 2309, |
| "pixel_diff_ratio": 0.0011135223765432098 |
| } |
| }</pre></details> |
| </article> |
| <article class="card "> |
| <div><span class="badge">NEGATIVE</span> <span class="badge valid">valid diff</span></div> |
| <div class="meta"><span>wrong_navigation</span><span>diff 0.09%</span><span>s00_neg_cd_src</span></div> |
| <img src="assets/step00_traj0_r_s00_neg_cd_src_s000213.png" alt="next state s00_neg_cd_src"> |
| <div class="cmd">pyautogui.write(message='cd src') |
| pyautogui.press('enter')</div> |
| <details><summary>OpenCUA row JSON</summary><pre class="json">{ |
| "images": [ |
| "/NHNHOME/data/sanghyeok/qwen-cua/data/tac_gui_traj_branch_v1/images/step00_traj0_r_current_s000000.png" |
| ], |
| "wm_next_image": "/NHNHOME/data/sanghyeok/qwen-cua/data/tac_gui_traj_branch_v1/images/step00_traj0_r_s00_neg_cd_src_s000213.png", |
| "messages": [ |
| { |
| "role": "system", |
| "content": "You are a GUI agent operating a desktop. For each step, given the user task and the current screenshot, respond in this exact format:\n\n## Thought:\n<brief reasoning about progress and the chosen next local action>\n\n## Action:\n<one-sentence description of what to do>\n\n## Code:\n<pyautogui code; coordinates normalized to [0,1] when clicking>" |
| }, |
| { |
| "role": "user", |
| "content": "Task: In the Linux terminal, inspect docs/guide.txt, return to the root folder, then enter the data folder.\n<image>\nWhat is the next action?" |
| }, |
| { |
| "role": "assistant", |
| "content": "## Thought:\nThis is a plausible but incorrect local action for the same state: 'cd src'.\n\n## Action:\nType the command 'cd src' and press Enter.\n\n## Code:\npyautogui.write(message='cd src')\npyautogui.press('enter')" |
| } |
| ], |
| "meta": { |
| "source": "tac_gui_traj_branch_v1", |
| "base_source": "cf_terminal_traj_fhd", |
| "trajectory_id": "tac_terminal_docs_data_v1", |
| "state_key": "tac_terminal_docs_data_v1/step0", |
| "base_state_key": "traj0/r", |
| "step_index": 0, |
| "task": "In the Linux terminal, inspect docs/guide.txt, return to the root folder, then enter the data folder.", |
| "step_intent": "Open the docs folder first.", |
| "branch_id": "s00_neg_cd_src", |
| "branch_idx": 2, |
| "branch_type": "wrong_navigation", |
| "is_gold": false, |
| "gold_command": "cd docs", |
| "command": "cd src", |
| "action_granularity": "policy_step_local_action", |
| "code_norm": "pyautogui.write(message='cd src')\npyautogui.press('enter')", |
| "current_image_rel": "images/step00_traj0_r_current_s000000.png", |
| "next_image_rel": "images/step00_traj0_r_s00_neg_cd_src_s000213.png", |
| "pixel_diff_nonzero": 1816, |
| "pixel_diff_ratio": 0.0008757716049382716 |
| } |
| }</pre></details> |
| </article> |
| <article class="card "> |
| <div><span class="badge">NEGATIVE</span> <span class="badge valid">valid diff</span></div> |
| <div class="meta"><span>wrong_file_read</span><span>diff 0.18%</span><span>s00_neg_cat_notes_md</span></div> |
| <img src="assets/step00_traj0_r_s00_neg_cat_notes_md_s000319.png" alt="next state s00_neg_cat_notes_md"> |
| <div class="cmd">pyautogui.write(message='cat notes.md') |
| pyautogui.press('enter')</div> |
| <details><summary>OpenCUA row JSON</summary><pre class="json">{ |
| "images": [ |
| "/NHNHOME/data/sanghyeok/qwen-cua/data/tac_gui_traj_branch_v1/images/step00_traj0_r_current_s000000.png" |
| ], |
| "wm_next_image": "/NHNHOME/data/sanghyeok/qwen-cua/data/tac_gui_traj_branch_v1/images/step00_traj0_r_s00_neg_cat_notes_md_s000319.png", |
| "messages": [ |
| { |
| "role": "system", |
| "content": "You are a GUI agent operating a desktop. For each step, given the user task and the current screenshot, respond in this exact format:\n\n## Thought:\n<brief reasoning about progress and the chosen next local action>\n\n## Action:\n<one-sentence description of what to do>\n\n## Code:\n<pyautogui code; coordinates normalized to [0,1] when clicking>" |
| }, |
| { |
| "role": "user", |
| "content": "Task: In the Linux terminal, inspect docs/guide.txt, return to the root folder, then enter the data folder.\n<image>\nWhat is the next action?" |
| }, |
| { |
| "role": "assistant", |
| "content": "## Thought:\nThis is a plausible but incorrect local action for the same state: 'cat notes.md'.\n\n## Action:\nType the command 'cat notes.md' and press Enter.\n\n## Code:\npyautogui.write(message='cat notes.md')\npyautogui.press('enter')" |
| } |
| ], |
| "meta": { |
| "source": "tac_gui_traj_branch_v1", |
| "base_source": "cf_terminal_traj_fhd", |
| "trajectory_id": "tac_terminal_docs_data_v1", |
| "state_key": "tac_terminal_docs_data_v1/step0", |
| "base_state_key": "traj0/r", |
| "step_index": 0, |
| "task": "In the Linux terminal, inspect docs/guide.txt, return to the root folder, then enter the data folder.", |
| "step_intent": "Open the docs folder first.", |
| "branch_id": "s00_neg_cat_notes_md", |
| "branch_idx": 3, |
| "branch_type": "wrong_file_read", |
| "is_gold": false, |
| "gold_command": "cd docs", |
| "command": "cat notes.md", |
| "action_granularity": "policy_step_local_action", |
| "code_norm": "pyautogui.write(message='cat notes.md')\npyautogui.press('enter')", |
| "current_image_rel": "images/step00_traj0_r_current_s000000.png", |
| "next_image_rel": "images/step00_traj0_r_s00_neg_cat_notes_md_s000319.png", |
| "pixel_diff_nonzero": 3726, |
| "pixel_diff_ratio": 0.001796875 |
| } |
| }</pre></details> |
| </article></div></div> |
| </div> |
| </section> |
| |
| <section class="step" id="step-1"> |
| <div class="step-head"> |
| <div> |
| <div class="eyebrow">Step 1 · same current state, multiple candidate local actions</div> |
| <div class="intent">Inspect the requested guide file.</div> |
| <div class="path">base_state=traj0/r.1 · gold=cat guide.txt</div> |
| </div> |
| </div> |
| <div class="grid"> |
| <div class="current"><h3>Current state s_t</h3><img src="assets/step01_traj0_r_1_current_s000108.png" alt="current step 1"></div> |
| <div class="branches"><h3>Candidate branches: s_t + action -> s_t+1</h3><div class="branch-grid"> |
| <article class="card gold"> |
| <div><span class="badge gold">GOLD</span> <span class="badge valid">valid diff</span></div> |
| <div class="meta"><span>gold</span><span>diff 0.22%</span><span>s01_gold_cat_guide_txt</span></div> |
| <img src="assets/step01_traj0_r_1_s01_gold_cat_guide_txt_s000161.png" alt="next state s01_gold_cat_guide_txt"> |
| <div class="cmd">pyautogui.write(message='cat guide.txt') |
| pyautogui.press('enter')</div> |
| <details><summary>OpenCUA row JSON</summary><pre class="json">{ |
| "images": [ |
| "/NHNHOME/data/sanghyeok/qwen-cua/data/tac_gui_traj_branch_v1/images/step01_traj0_r_1_current_s000108.png" |
| ], |
| "wm_next_image": "/NHNHOME/data/sanghyeok/qwen-cua/data/tac_gui_traj_branch_v1/images/step01_traj0_r_1_s01_gold_cat_guide_txt_s000161.png", |
| "messages": [ |
| { |
| "role": "system", |
| "content": "You are a GUI agent operating a desktop. For each step, given the user task and the current screenshot, respond in this exact format:\n\n## Thought:\n<brief reasoning about progress and the chosen next local action>\n\n## Action:\n<one-sentence description of what to do>\n\n## Code:\n<pyautogui code; coordinates normalized to [0,1] when clicking>" |
| }, |
| { |
| "role": "user", |
| "content": "<image>\nWhat is the next action?" |
| }, |
| { |
| "role": "assistant", |
| "content": "## Thought:\nProgress: Open the docs folder first. The next local terminal action should be 'cd docs'.\n\n## Action:\nType the command 'cd docs' and press Enter.\n\n## Code:\npyautogui.write(message='cd docs')\npyautogui.press('enter')" |
| }, |
| { |
| "role": "user", |
| "content": "Task: In the Linux terminal, inspect docs/guide.txt, return to the root folder, then enter the data folder.\n<image>\nWhat is the next action?" |
| }, |
| { |
| "role": "assistant", |
| "content": "## Thought:\nProgress: Inspect the requested guide file. The next local terminal action should be 'cat guide.txt'.\n\n## Action:\nType the command 'cat guide.txt' and press Enter.\n\n## Code:\npyautogui.write(message='cat guide.txt')\npyautogui.press('enter')" |
| } |
| ], |
| "meta": { |
| "source": "tac_gui_traj_branch_v1", |
| "base_source": "cf_terminal_traj_fhd", |
| "trajectory_id": "tac_terminal_docs_data_v1", |
| "state_key": "tac_terminal_docs_data_v1/step1", |
| "base_state_key": "traj0/r.1", |
| "step_index": 1, |
| "task": "In the Linux terminal, inspect docs/guide.txt, return to the root folder, then enter the data folder.", |
| "step_intent": "Inspect the requested guide file.", |
| "branch_id": "s01_gold_cat_guide_txt", |
| "branch_idx": 2, |
| "branch_type": "gold", |
| "is_gold": true, |
| "gold_command": "cat guide.txt", |
| "command": "cat guide.txt", |
| "action_granularity": "policy_step_local_action", |
| "code_norm": "pyautogui.write(message='cat guide.txt')\npyautogui.press('enter')", |
| "current_image_rel": "images/step01_traj0_r_1_current_s000108.png", |
| "next_image_rel": "images/step01_traj0_r_1_s01_gold_cat_guide_txt_s000161.png", |
| "pixel_diff_nonzero": 4533, |
| "pixel_diff_ratio": 0.0021860532407407406 |
| } |
| }</pre></details> |
| </article> |
| <article class="card "> |
| <div><span class="badge">NEGATIVE</span> <span class="badge valid">valid diff</span></div> |
| <div class="meta"><span>wrong_navigation</span><span>diff 0.12%</span><span>s01_neg_cd</span></div> |
| <img src="assets/step01_traj0_r_1_s01_neg_cd_s000109.png" alt="next state s01_neg_cd"> |
| <div class="cmd">pyautogui.write(message='cd ..') |
| pyautogui.press('enter')</div> |
| <details><summary>OpenCUA row JSON</summary><pre class="json">{ |
| "images": [ |
| "/NHNHOME/data/sanghyeok/qwen-cua/data/tac_gui_traj_branch_v1/images/step01_traj0_r_1_current_s000108.png" |
| ], |
| "wm_next_image": "/NHNHOME/data/sanghyeok/qwen-cua/data/tac_gui_traj_branch_v1/images/step01_traj0_r_1_s01_neg_cd_s000109.png", |
| "messages": [ |
| { |
| "role": "system", |
| "content": "You are a GUI agent operating a desktop. For each step, given the user task and the current screenshot, respond in this exact format:\n\n## Thought:\n<brief reasoning about progress and the chosen next local action>\n\n## Action:\n<one-sentence description of what to do>\n\n## Code:\n<pyautogui code; coordinates normalized to [0,1] when clicking>" |
| }, |
| { |
| "role": "user", |
| "content": "<image>\nWhat is the next action?" |
| }, |
| { |
| "role": "assistant", |
| "content": "## Thought:\nProgress: Open the docs folder first. The next local terminal action should be 'cd docs'.\n\n## Action:\nType the command 'cd docs' and press Enter.\n\n## Code:\npyautogui.write(message='cd docs')\npyautogui.press('enter')" |
| }, |
| { |
| "role": "user", |
| "content": "Task: In the Linux terminal, inspect docs/guide.txt, return to the root folder, then enter the data folder.\n<image>\nWhat is the next action?" |
| }, |
| { |
| "role": "assistant", |
| "content": "## Thought:\nThis is a plausible but incorrect local action for the same state: 'cd ..'.\n\n## Action:\nType the command 'cd ..' and press Enter.\n\n## Code:\npyautogui.write(message='cd ..')\npyautogui.press('enter')" |
| } |
| ], |
| "meta": { |
| "source": "tac_gui_traj_branch_v1", |
| "base_source": "cf_terminal_traj_fhd", |
| "trajectory_id": "tac_terminal_docs_data_v1", |
| "state_key": "tac_terminal_docs_data_v1/step1", |
| "base_state_key": "traj0/r.1", |
| "step_index": 1, |
| "task": "In the Linux terminal, inspect docs/guide.txt, return to the root folder, then enter the data folder.", |
| "step_intent": "Inspect the requested guide file.", |
| "branch_id": "s01_neg_cd", |
| "branch_idx": 0, |
| "branch_type": "wrong_navigation", |
| "is_gold": false, |
| "gold_command": "cat guide.txt", |
| "command": "cd ..", |
| "action_granularity": "policy_step_local_action", |
| "code_norm": "pyautogui.write(message='cd ..')\npyautogui.press('enter')", |
| "current_image_rel": "images/step01_traj0_r_1_current_s000108.png", |
| "next_image_rel": "images/step01_traj0_r_1_s01_neg_cd_s000109.png", |
| "pixel_diff_nonzero": 2582, |
| "pixel_diff_ratio": 0.0012451774691358026 |
| } |
| }</pre></details> |
| </article> |
| <article class="card "> |
| <div><span class="badge">NEGATIVE</span> <span class="badge valid">valid diff</span></div> |
| <div class="meta"><span>wrong_file_read</span><span>diff 0.16%</span><span>s01_neg_cat_faq_txt</span></div> |
| <img src="assets/step01_traj0_r_1_s01_neg_cat_faq_txt_s000135.png" alt="next state s01_neg_cat_faq_txt"> |
| <div class="cmd">pyautogui.write(message='cat faq.txt') |
| pyautogui.press('enter')</div> |
| <details><summary>OpenCUA row JSON</summary><pre class="json">{ |
| "images": [ |
| "/NHNHOME/data/sanghyeok/qwen-cua/data/tac_gui_traj_branch_v1/images/step01_traj0_r_1_current_s000108.png" |
| ], |
| "wm_next_image": "/NHNHOME/data/sanghyeok/qwen-cua/data/tac_gui_traj_branch_v1/images/step01_traj0_r_1_s01_neg_cat_faq_txt_s000135.png", |
| "messages": [ |
| { |
| "role": "system", |
| "content": "You are a GUI agent operating a desktop. For each step, given the user task and the current screenshot, respond in this exact format:\n\n## Thought:\n<brief reasoning about progress and the chosen next local action>\n\n## Action:\n<one-sentence description of what to do>\n\n## Code:\n<pyautogui code; coordinates normalized to [0,1] when clicking>" |
| }, |
| { |
| "role": "user", |
| "content": "<image>\nWhat is the next action?" |
| }, |
| { |
| "role": "assistant", |
| "content": "## Thought:\nProgress: Open the docs folder first. The next local terminal action should be 'cd docs'.\n\n## Action:\nType the command 'cd docs' and press Enter.\n\n## Code:\npyautogui.write(message='cd docs')\npyautogui.press('enter')" |
| }, |
| { |
| "role": "user", |
| "content": "Task: In the Linux terminal, inspect docs/guide.txt, return to the root folder, then enter the data folder.\n<image>\nWhat is the next action?" |
| }, |
| { |
| "role": "assistant", |
| "content": "## Thought:\nThis is a plausible but incorrect local action for the same state: 'cat faq.txt'.\n\n## Action:\nType the command 'cat faq.txt' and press Enter.\n\n## Code:\npyautogui.write(message='cat faq.txt')\npyautogui.press('enter')" |
| } |
| ], |
| "meta": { |
| "source": "tac_gui_traj_branch_v1", |
| "base_source": "cf_terminal_traj_fhd", |
| "trajectory_id": "tac_terminal_docs_data_v1", |
| "state_key": "tac_terminal_docs_data_v1/step1", |
| "base_state_key": "traj0/r.1", |
| "step_index": 1, |
| "task": "In the Linux terminal, inspect docs/guide.txt, return to the root folder, then enter the data folder.", |
| "step_intent": "Inspect the requested guide file.", |
| "branch_id": "s01_neg_cat_faq_txt", |
| "branch_idx": 1, |
| "branch_type": "wrong_file_read", |
| "is_gold": false, |
| "gold_command": "cat guide.txt", |
| "command": "cat faq.txt", |
| "action_granularity": "policy_step_local_action", |
| "code_norm": "pyautogui.write(message='cat faq.txt')\npyautogui.press('enter')", |
| "current_image_rel": "images/step01_traj0_r_1_current_s000108.png", |
| "next_image_rel": "images/step01_traj0_r_1_s01_neg_cat_faq_txt_s000135.png", |
| "pixel_diff_nonzero": 3354, |
| "pixel_diff_ratio": 0.001617476851851852 |
| } |
| }</pre></details> |
| </article> |
| <article class="card "> |
| <div><span class="badge">NEGATIVE</span> <span class="badge valid">valid diff</span></div> |
| <div class="meta"><span>wrong_inspection</span><span>diff 0.37%</span><span>s01_neg_ls_l</span></div> |
| <img src="assets/step01_traj0_r_1_s01_neg_ls_l_s000187.png" alt="next state s01_neg_ls_l"> |
| <div class="cmd">pyautogui.write(message='ls -l') |
| pyautogui.press('enter')</div> |
| <details><summary>OpenCUA row JSON</summary><pre class="json">{ |
| "images": [ |
| "/NHNHOME/data/sanghyeok/qwen-cua/data/tac_gui_traj_branch_v1/images/step01_traj0_r_1_current_s000108.png" |
| ], |
| "wm_next_image": "/NHNHOME/data/sanghyeok/qwen-cua/data/tac_gui_traj_branch_v1/images/step01_traj0_r_1_s01_neg_ls_l_s000187.png", |
| "messages": [ |
| { |
| "role": "system", |
| "content": "You are a GUI agent operating a desktop. For each step, given the user task and the current screenshot, respond in this exact format:\n\n## Thought:\n<brief reasoning about progress and the chosen next local action>\n\n## Action:\n<one-sentence description of what to do>\n\n## Code:\n<pyautogui code; coordinates normalized to [0,1] when clicking>" |
| }, |
| { |
| "role": "user", |
| "content": "<image>\nWhat is the next action?" |
| }, |
| { |
| "role": "assistant", |
| "content": "## Thought:\nProgress: Open the docs folder first. The next local terminal action should be 'cd docs'.\n\n## Action:\nType the command 'cd docs' and press Enter.\n\n## Code:\npyautogui.write(message='cd docs')\npyautogui.press('enter')" |
| }, |
| { |
| "role": "user", |
| "content": "Task: In the Linux terminal, inspect docs/guide.txt, return to the root folder, then enter the data folder.\n<image>\nWhat is the next action?" |
| }, |
| { |
| "role": "assistant", |
| "content": "## Thought:\nThis is a plausible but incorrect local action for the same state: 'ls -l'.\n\n## Action:\nType the command 'ls -l' and press Enter.\n\n## Code:\npyautogui.write(message='ls -l')\npyautogui.press('enter')" |
| } |
| ], |
| "meta": { |
| "source": "tac_gui_traj_branch_v1", |
| "base_source": "cf_terminal_traj_fhd", |
| "trajectory_id": "tac_terminal_docs_data_v1", |
| "state_key": "tac_terminal_docs_data_v1/step1", |
| "base_state_key": "traj0/r.1", |
| "step_index": 1, |
| "task": "In the Linux terminal, inspect docs/guide.txt, return to the root folder, then enter the data folder.", |
| "step_intent": "Inspect the requested guide file.", |
| "branch_id": "s01_neg_ls_l", |
| "branch_idx": 3, |
| "branch_type": "wrong_inspection", |
| "is_gold": false, |
| "gold_command": "cat guide.txt", |
| "command": "ls -l", |
| "action_granularity": "policy_step_local_action", |
| "code_norm": "pyautogui.write(message='ls -l')\npyautogui.press('enter')", |
| "current_image_rel": "images/step01_traj0_r_1_current_s000108.png", |
| "next_image_rel": "images/step01_traj0_r_1_s01_neg_ls_l_s000187.png", |
| "pixel_diff_nonzero": 7599, |
| "pixel_diff_ratio": 0.003664641203703704 |
| } |
| }</pre></details> |
| </article></div></div> |
| </div> |
| </section> |
| |
| <section class="step" id="step-2"> |
| <div class="step-head"> |
| <div> |
| <div class="eyebrow">Step 2 · same current state, multiple candidate local actions</div> |
| <div class="intent">Return from docs to the root folder.</div> |
| <div class="path">base_state=traj0/r.1.2 · gold=cd ..</div> |
| </div> |
| </div> |
| <div class="grid"> |
| <div class="current"><h3>Current state s_t</h3><img src="assets/step02_traj0_r_1_2_current_s000162.png" alt="current step 2"></div> |
| <div class="branches"><h3>Candidate branches: s_t + action -> s_t+1</h3><div class="branch-grid"> |
| <article class="card gold"> |
| <div><span class="badge gold">GOLD</span> <span class="badge valid">valid diff</span></div> |
| <div class="meta"><span>gold</span><span>diff 0.20%</span><span>s02_gold_cd</span></div> |
| <img src="assets/step02_traj0_r_1_2_s02_gold_cd_s000163.png" alt="next state s02_gold_cd"> |
| <div class="cmd">pyautogui.write(message='cd ..') |
| pyautogui.press('enter')</div> |
| <details><summary>OpenCUA row JSON</summary><pre class="json">{ |
| "images": [ |
| "/NHNHOME/data/sanghyeok/qwen-cua/data/tac_gui_traj_branch_v1/images/step02_traj0_r_1_2_current_s000162.png" |
| ], |
| "wm_next_image": "/NHNHOME/data/sanghyeok/qwen-cua/data/tac_gui_traj_branch_v1/images/step02_traj0_r_1_2_s02_gold_cd_s000163.png", |
| "messages": [ |
| { |
| "role": "system", |
| "content": "You are a GUI agent operating a desktop. For each step, given the user task and the current screenshot, respond in this exact format:\n\n## Thought:\n<brief reasoning about progress and the chosen next local action>\n\n## Action:\n<one-sentence description of what to do>\n\n## Code:\n<pyautogui code; coordinates normalized to [0,1] when clicking>" |
| }, |
| { |
| "role": "user", |
| "content": "<image>\nWhat is the next action?" |
| }, |
| { |
| "role": "assistant", |
| "content": "## Thought:\nProgress: Open the docs folder first. The next local terminal action should be 'cd docs'.\n\n## Action:\nType the command 'cd docs' and press Enter.\n\n## Code:\npyautogui.write(message='cd docs')\npyautogui.press('enter')" |
| }, |
| { |
| "role": "user", |
| "content": "<image>\nWhat is the next action?" |
| }, |
| { |
| "role": "assistant", |
| "content": "## Thought:\nProgress: Inspect the requested guide file. The next local terminal action should be 'cat guide.txt'.\n\n## Action:\nType the command 'cat guide.txt' and press Enter.\n\n## Code:\npyautogui.write(message='cat guide.txt')\npyautogui.press('enter')" |
| }, |
| { |
| "role": "user", |
| "content": "Task: In the Linux terminal, inspect docs/guide.txt, return to the root folder, then enter the data folder.\n<image>\nWhat is the next action?" |
| }, |
| { |
| "role": "assistant", |
| "content": "## Thought:\nProgress: Return from docs to the root folder. The next local terminal action should be 'cd ..'.\n\n## Action:\nType the command 'cd ..' and press Enter.\n\n## Code:\npyautogui.write(message='cd ..')\npyautogui.press('enter')" |
| } |
| ], |
| "meta": { |
| "source": "tac_gui_traj_branch_v1", |
| "base_source": "cf_terminal_traj_fhd", |
| "trajectory_id": "tac_terminal_docs_data_v1", |
| "state_key": "tac_terminal_docs_data_v1/step2", |
| "base_state_key": "traj0/r.1.2", |
| "step_index": 2, |
| "task": "In the Linux terminal, inspect docs/guide.txt, return to the root folder, then enter the data folder.", |
| "step_intent": "Return from docs to the root folder.", |
| "branch_id": "s02_gold_cd", |
| "branch_idx": 0, |
| "branch_type": "gold", |
| "is_gold": true, |
| "gold_command": "cd ..", |
| "command": "cd ..", |
| "action_granularity": "policy_step_local_action", |
| "code_norm": "pyautogui.write(message='cd ..')\npyautogui.press('enter')", |
| "current_image_rel": "images/step02_traj0_r_1_2_current_s000162.png", |
| "next_image_rel": "images/step02_traj0_r_1_2_s02_gold_cd_s000163.png", |
| "pixel_diff_nonzero": 4187, |
| "pixel_diff_ratio": 0.002019193672839506 |
| } |
| }</pre></details> |
| </article> |
| <article class="card "> |
| <div><span class="badge">NEGATIVE</span> <span class="badge valid">valid diff</span></div> |
| <div class="meta"><span>wrong_file_read</span><span>diff 0.16%</span><span>s02_neg_cat_faq_txt</span></div> |
| <img src="assets/step02_traj0_r_1_2_s02_neg_cat_faq_txt_s000169.png" alt="next state s02_neg_cat_faq_txt"> |
| <div class="cmd">pyautogui.write(message='cat faq.txt') |
| pyautogui.press('enter')</div> |
| <details><summary>OpenCUA row JSON</summary><pre class="json">{ |
| "images": [ |
| "/NHNHOME/data/sanghyeok/qwen-cua/data/tac_gui_traj_branch_v1/images/step02_traj0_r_1_2_current_s000162.png" |
| ], |
| "wm_next_image": "/NHNHOME/data/sanghyeok/qwen-cua/data/tac_gui_traj_branch_v1/images/step02_traj0_r_1_2_s02_neg_cat_faq_txt_s000169.png", |
| "messages": [ |
| { |
| "role": "system", |
| "content": "You are a GUI agent operating a desktop. For each step, given the user task and the current screenshot, respond in this exact format:\n\n## Thought:\n<brief reasoning about progress and the chosen next local action>\n\n## Action:\n<one-sentence description of what to do>\n\n## Code:\n<pyautogui code; coordinates normalized to [0,1] when clicking>" |
| }, |
| { |
| "role": "user", |
| "content": "<image>\nWhat is the next action?" |
| }, |
| { |
| "role": "assistant", |
| "content": "## Thought:\nProgress: Open the docs folder first. The next local terminal action should be 'cd docs'.\n\n## Action:\nType the command 'cd docs' and press Enter.\n\n## Code:\npyautogui.write(message='cd docs')\npyautogui.press('enter')" |
| }, |
| { |
| "role": "user", |
| "content": "<image>\nWhat is the next action?" |
| }, |
| { |
| "role": "assistant", |
| "content": "## Thought:\nProgress: Inspect the requested guide file. The next local terminal action should be 'cat guide.txt'.\n\n## Action:\nType the command 'cat guide.txt' and press Enter.\n\n## Code:\npyautogui.write(message='cat guide.txt')\npyautogui.press('enter')" |
| }, |
| { |
| "role": "user", |
| "content": "Task: In the Linux terminal, inspect docs/guide.txt, return to the root folder, then enter the data folder.\n<image>\nWhat is the next action?" |
| }, |
| { |
| "role": "assistant", |
| "content": "## Thought:\nThis is a plausible but incorrect local action for the same state: 'cat faq.txt'.\n\n## Action:\nType the command 'cat faq.txt' and press Enter.\n\n## Code:\npyautogui.write(message='cat faq.txt')\npyautogui.press('enter')" |
| } |
| ], |
| "meta": { |
| "source": "tac_gui_traj_branch_v1", |
| "base_source": "cf_terminal_traj_fhd", |
| "trajectory_id": "tac_terminal_docs_data_v1", |
| "state_key": "tac_terminal_docs_data_v1/step2", |
| "base_state_key": "traj0/r.1.2", |
| "step_index": 2, |
| "task": "In the Linux terminal, inspect docs/guide.txt, return to the root folder, then enter the data folder.", |
| "step_intent": "Return from docs to the root folder.", |
| "branch_id": "s02_neg_cat_faq_txt", |
| "branch_idx": 1, |
| "branch_type": "wrong_file_read", |
| "is_gold": false, |
| "gold_command": "cd ..", |
| "command": "cat faq.txt", |
| "action_granularity": "policy_step_local_action", |
| "code_norm": "pyautogui.write(message='cat faq.txt')\npyautogui.press('enter')", |
| "current_image_rel": "images/step02_traj0_r_1_2_current_s000162.png", |
| "next_image_rel": "images/step02_traj0_r_1_2_s02_neg_cat_faq_txt_s000169.png", |
| "pixel_diff_nonzero": 3354, |
| "pixel_diff_ratio": 0.001617476851851852 |
| } |
| }</pre></details> |
| </article> |
| <article class="card "> |
| <div><span class="badge">NEGATIVE</span> <span class="badge valid">valid diff</span></div> |
| <div class="meta"><span>wrong_file_read</span><span>diff 0.22%</span><span>s02_neg_cat_guide_txt</span></div> |
| <img src="assets/step02_traj0_r_1_2_s02_neg_cat_guide_txt_s000175.png" alt="next state s02_neg_cat_guide_txt"> |
| <div class="cmd">pyautogui.write(message='cat guide.txt') |
| pyautogui.press('enter')</div> |
| <details><summary>OpenCUA row JSON</summary><pre class="json">{ |
| "images": [ |
| "/NHNHOME/data/sanghyeok/qwen-cua/data/tac_gui_traj_branch_v1/images/step02_traj0_r_1_2_current_s000162.png" |
| ], |
| "wm_next_image": "/NHNHOME/data/sanghyeok/qwen-cua/data/tac_gui_traj_branch_v1/images/step02_traj0_r_1_2_s02_neg_cat_guide_txt_s000175.png", |
| "messages": [ |
| { |
| "role": "system", |
| "content": "You are a GUI agent operating a desktop. For each step, given the user task and the current screenshot, respond in this exact format:\n\n## Thought:\n<brief reasoning about progress and the chosen next local action>\n\n## Action:\n<one-sentence description of what to do>\n\n## Code:\n<pyautogui code; coordinates normalized to [0,1] when clicking>" |
| }, |
| { |
| "role": "user", |
| "content": "<image>\nWhat is the next action?" |
| }, |
| { |
| "role": "assistant", |
| "content": "## Thought:\nProgress: Open the docs folder first. The next local terminal action should be 'cd docs'.\n\n## Action:\nType the command 'cd docs' and press Enter.\n\n## Code:\npyautogui.write(message='cd docs')\npyautogui.press('enter')" |
| }, |
| { |
| "role": "user", |
| "content": "<image>\nWhat is the next action?" |
| }, |
| { |
| "role": "assistant", |
| "content": "## Thought:\nProgress: Inspect the requested guide file. The next local terminal action should be 'cat guide.txt'.\n\n## Action:\nType the command 'cat guide.txt' and press Enter.\n\n## Code:\npyautogui.write(message='cat guide.txt')\npyautogui.press('enter')" |
| }, |
| { |
| "role": "user", |
| "content": "Task: In the Linux terminal, inspect docs/guide.txt, return to the root folder, then enter the data folder.\n<image>\nWhat is the next action?" |
| }, |
| { |
| "role": "assistant", |
| "content": "## Thought:\nThis is a plausible but incorrect local action for the same state: 'cat guide.txt'.\n\n## Action:\nType the command 'cat guide.txt' and press Enter.\n\n## Code:\npyautogui.write(message='cat guide.txt')\npyautogui.press('enter')" |
| } |
| ], |
| "meta": { |
| "source": "tac_gui_traj_branch_v1", |
| "base_source": "cf_terminal_traj_fhd", |
| "trajectory_id": "tac_terminal_docs_data_v1", |
| "state_key": "tac_terminal_docs_data_v1/step2", |
| "base_state_key": "traj0/r.1.2", |
| "step_index": 2, |
| "task": "In the Linux terminal, inspect docs/guide.txt, return to the root folder, then enter the data folder.", |
| "step_intent": "Return from docs to the root folder.", |
| "branch_id": "s02_neg_cat_guide_txt", |
| "branch_idx": 2, |
| "branch_type": "wrong_file_read", |
| "is_gold": false, |
| "gold_command": "cd ..", |
| "command": "cat guide.txt", |
| "action_granularity": "policy_step_local_action", |
| "code_norm": "pyautogui.write(message='cat guide.txt')\npyautogui.press('enter')", |
| "current_image_rel": "images/step02_traj0_r_1_2_current_s000162.png", |
| "next_image_rel": "images/step02_traj0_r_1_2_s02_neg_cat_guide_txt_s000175.png", |
| "pixel_diff_nonzero": 4533, |
| "pixel_diff_ratio": 0.0021860532407407406 |
| } |
| }</pre></details> |
| </article> |
| <article class="card "> |
| <div><span class="badge">NEGATIVE</span> <span class="badge valid">valid diff</span></div> |
| <div class="meta"><span>wrong_inspection</span><span>diff 0.37%</span><span>s02_neg_ls_l</span></div> |
| <img src="assets/step02_traj0_r_1_2_s02_neg_ls_l_s000181.png" alt="next state s02_neg_ls_l"> |
| <div class="cmd">pyautogui.write(message='ls -l') |
| pyautogui.press('enter')</div> |
| <details><summary>OpenCUA row JSON</summary><pre class="json">{ |
| "images": [ |
| "/NHNHOME/data/sanghyeok/qwen-cua/data/tac_gui_traj_branch_v1/images/step02_traj0_r_1_2_current_s000162.png" |
| ], |
| "wm_next_image": "/NHNHOME/data/sanghyeok/qwen-cua/data/tac_gui_traj_branch_v1/images/step02_traj0_r_1_2_s02_neg_ls_l_s000181.png", |
| "messages": [ |
| { |
| "role": "system", |
| "content": "You are a GUI agent operating a desktop. For each step, given the user task and the current screenshot, respond in this exact format:\n\n## Thought:\n<brief reasoning about progress and the chosen next local action>\n\n## Action:\n<one-sentence description of what to do>\n\n## Code:\n<pyautogui code; coordinates normalized to [0,1] when clicking>" |
| }, |
| { |
| "role": "user", |
| "content": "<image>\nWhat is the next action?" |
| }, |
| { |
| "role": "assistant", |
| "content": "## Thought:\nProgress: Open the docs folder first. The next local terminal action should be 'cd docs'.\n\n## Action:\nType the command 'cd docs' and press Enter.\n\n## Code:\npyautogui.write(message='cd docs')\npyautogui.press('enter')" |
| }, |
| { |
| "role": "user", |
| "content": "<image>\nWhat is the next action?" |
| }, |
| { |
| "role": "assistant", |
| "content": "## Thought:\nProgress: Inspect the requested guide file. The next local terminal action should be 'cat guide.txt'.\n\n## Action:\nType the command 'cat guide.txt' and press Enter.\n\n## Code:\npyautogui.write(message='cat guide.txt')\npyautogui.press('enter')" |
| }, |
| { |
| "role": "user", |
| "content": "Task: In the Linux terminal, inspect docs/guide.txt, return to the root folder, then enter the data folder.\n<image>\nWhat is the next action?" |
| }, |
| { |
| "role": "assistant", |
| "content": "## Thought:\nThis is a plausible but incorrect local action for the same state: 'ls -l'.\n\n## Action:\nType the command 'ls -l' and press Enter.\n\n## Code:\npyautogui.write(message='ls -l')\npyautogui.press('enter')" |
| } |
| ], |
| "meta": { |
| "source": "tac_gui_traj_branch_v1", |
| "base_source": "cf_terminal_traj_fhd", |
| "trajectory_id": "tac_terminal_docs_data_v1", |
| "state_key": "tac_terminal_docs_data_v1/step2", |
| "base_state_key": "traj0/r.1.2", |
| "step_index": 2, |
| "task": "In the Linux terminal, inspect docs/guide.txt, return to the root folder, then enter the data folder.", |
| "step_intent": "Return from docs to the root folder.", |
| "branch_id": "s02_neg_ls_l", |
| "branch_idx": 3, |
| "branch_type": "wrong_inspection", |
| "is_gold": false, |
| "gold_command": "cd ..", |
| "command": "ls -l", |
| "action_granularity": "policy_step_local_action", |
| "code_norm": "pyautogui.write(message='ls -l')\npyautogui.press('enter')", |
| "current_image_rel": "images/step02_traj0_r_1_2_current_s000162.png", |
| "next_image_rel": "images/step02_traj0_r_1_2_s02_neg_ls_l_s000181.png", |
| "pixel_diff_nonzero": 7599, |
| "pixel_diff_ratio": 0.003664641203703704 |
| } |
| }</pre></details> |
| </article></div></div> |
| </div> |
| </section> |
| |
| <section class="step" id="step-3"> |
| <div class="step-head"> |
| <div> |
| <div class="eyebrow">Step 3 · same current state, multiple candidate local actions</div> |
| <div class="intent">Enter the data folder after returning to root.</div> |
| <div class="path">base_state=traj0/r.1.2.0 · gold=cd data</div> |
| </div> |
| </div> |
| <div class="grid"> |
| <div class="current"><h3>Current state s_t</h3><img src="assets/step03_traj0_r_1_2_0_current_s000164.png" alt="current step 3"></div> |
| <div class="branches"><h3>Candidate branches: s_t + action -> s_t+1</h3><div class="branch-grid"> |
| <article class="card gold"> |
| <div><span class="badge gold">GOLD</span> <span class="badge valid">valid diff</span></div> |
| <div class="meta"><span>gold</span><span>diff 0.28%</span><span>s03_gold_cd_data</span></div> |
| <img src="assets/step03_traj0_r_1_2_0_s03_gold_cd_data_s000165.png" alt="next state s03_gold_cd_data"> |
| <div class="cmd">pyautogui.write(message='cd data') |
| pyautogui.press('enter')</div> |
| <details><summary>OpenCUA row JSON</summary><pre class="json">{ |
| "images": [ |
| "/NHNHOME/data/sanghyeok/qwen-cua/data/tac_gui_traj_branch_v1/images/step03_traj0_r_1_2_0_current_s000164.png" |
| ], |
| "wm_next_image": "/NHNHOME/data/sanghyeok/qwen-cua/data/tac_gui_traj_branch_v1/images/step03_traj0_r_1_2_0_s03_gold_cd_data_s000165.png", |
| "messages": [ |
| { |
| "role": "system", |
| "content": "You are a GUI agent operating a desktop. For each step, given the user task and the current screenshot, respond in this exact format:\n\n## Thought:\n<brief reasoning about progress and the chosen next local action>\n\n## Action:\n<one-sentence description of what to do>\n\n## Code:\n<pyautogui code; coordinates normalized to [0,1] when clicking>" |
| }, |
| { |
| "role": "user", |
| "content": "<image>\nWhat is the next action?" |
| }, |
| { |
| "role": "assistant", |
| "content": "## Thought:\nProgress: Inspect the requested guide file. The next local terminal action should be 'cat guide.txt'.\n\n## Action:\nType the command 'cat guide.txt' and press Enter.\n\n## Code:\npyautogui.write(message='cat guide.txt')\npyautogui.press('enter')" |
| }, |
| { |
| "role": "user", |
| "content": "<image>\nWhat is the next action?" |
| }, |
| { |
| "role": "assistant", |
| "content": "## Thought:\nProgress: Return from docs to the root folder. The next local terminal action should be 'cd ..'.\n\n## Action:\nType the command 'cd ..' and press Enter.\n\n## Code:\npyautogui.write(message='cd ..')\npyautogui.press('enter')" |
| }, |
| { |
| "role": "user", |
| "content": "Task: In the Linux terminal, inspect docs/guide.txt, return to the root folder, then enter the data folder.\n<image>\nWhat is the next action?" |
| }, |
| { |
| "role": "assistant", |
| "content": "## Thought:\nProgress: Enter the data folder after returning to root. The next local terminal action should be 'cd data'.\n\n## Action:\nType the command 'cd data' and press Enter.\n\n## Code:\npyautogui.write(message='cd data')\npyautogui.press('enter')" |
| } |
| ], |
| "meta": { |
| "source": "tac_gui_traj_branch_v1", |
| "base_source": "cf_terminal_traj_fhd", |
| "trajectory_id": "tac_terminal_docs_data_v1", |
| "state_key": "tac_terminal_docs_data_v1/step3", |
| "base_state_key": "traj0/r.1.2.0", |
| "step_index": 3, |
| "task": "In the Linux terminal, inspect docs/guide.txt, return to the root folder, then enter the data folder.", |
| "step_intent": "Enter the data folder after returning to root.", |
| "branch_id": "s03_gold_cd_data", |
| "branch_idx": 0, |
| "branch_type": "gold", |
| "is_gold": true, |
| "gold_command": "cd data", |
| "command": "cd data", |
| "action_granularity": "policy_step_local_action", |
| "code_norm": "pyautogui.write(message='cd data')\npyautogui.press('enter')", |
| "current_image_rel": "images/step03_traj0_r_1_2_0_current_s000164.png", |
| "next_image_rel": "images/step03_traj0_r_1_2_0_s03_gold_cd_data_s000165.png", |
| "pixel_diff_nonzero": 5746, |
| "pixel_diff_ratio": 0.0027710262345679013 |
| } |
| }</pre></details> |
| </article> |
| <article class="card "> |
| <div><span class="badge">NEGATIVE</span> <span class="badge valid">valid diff</span></div> |
| <div class="meta"><span>wrong_navigation</span><span>diff 0.27%</span><span>s03_neg_cd_docs</span></div> |
| <img src="assets/step03_traj0_r_1_2_0_s03_neg_cd_docs_s000166.png" alt="next state s03_neg_cd_docs"> |
| <div class="cmd">pyautogui.write(message='cd docs') |
| pyautogui.press('enter')</div> |
| <details><summary>OpenCUA row JSON</summary><pre class="json">{ |
| "images": [ |
| "/NHNHOME/data/sanghyeok/qwen-cua/data/tac_gui_traj_branch_v1/images/step03_traj0_r_1_2_0_current_s000164.png" |
| ], |
| "wm_next_image": "/NHNHOME/data/sanghyeok/qwen-cua/data/tac_gui_traj_branch_v1/images/step03_traj0_r_1_2_0_s03_neg_cd_docs_s000166.png", |
| "messages": [ |
| { |
| "role": "system", |
| "content": "You are a GUI agent operating a desktop. For each step, given the user task and the current screenshot, respond in this exact format:\n\n## Thought:\n<brief reasoning about progress and the chosen next local action>\n\n## Action:\n<one-sentence description of what to do>\n\n## Code:\n<pyautogui code; coordinates normalized to [0,1] when clicking>" |
| }, |
| { |
| "role": "user", |
| "content": "<image>\nWhat is the next action?" |
| }, |
| { |
| "role": "assistant", |
| "content": "## Thought:\nProgress: Inspect the requested guide file. The next local terminal action should be 'cat guide.txt'.\n\n## Action:\nType the command 'cat guide.txt' and press Enter.\n\n## Code:\npyautogui.write(message='cat guide.txt')\npyautogui.press('enter')" |
| }, |
| { |
| "role": "user", |
| "content": "<image>\nWhat is the next action?" |
| }, |
| { |
| "role": "assistant", |
| "content": "## Thought:\nProgress: Return from docs to the root folder. The next local terminal action should be 'cd ..'.\n\n## Action:\nType the command 'cd ..' and press Enter.\n\n## Code:\npyautogui.write(message='cd ..')\npyautogui.press('enter')" |
| }, |
| { |
| "role": "user", |
| "content": "Task: In the Linux terminal, inspect docs/guide.txt, return to the root folder, then enter the data folder.\n<image>\nWhat is the next action?" |
| }, |
| { |
| "role": "assistant", |
| "content": "## Thought:\nThis is a plausible but incorrect local action for the same state: 'cd docs'.\n\n## Action:\nType the command 'cd docs' and press Enter.\n\n## Code:\npyautogui.write(message='cd docs')\npyautogui.press('enter')" |
| } |
| ], |
| "meta": { |
| "source": "tac_gui_traj_branch_v1", |
| "base_source": "cf_terminal_traj_fhd", |
| "trajectory_id": "tac_terminal_docs_data_v1", |
| "state_key": "tac_terminal_docs_data_v1/step3", |
| "base_state_key": "traj0/r.1.2.0", |
| "step_index": 3, |
| "task": "In the Linux terminal, inspect docs/guide.txt, return to the root folder, then enter the data folder.", |
| "step_intent": "Enter the data folder after returning to root.", |
| "branch_id": "s03_neg_cd_docs", |
| "branch_idx": 1, |
| "branch_type": "wrong_navigation", |
| "is_gold": false, |
| "gold_command": "cd data", |
| "command": "cd docs", |
| "action_granularity": "policy_step_local_action", |
| "code_norm": "pyautogui.write(message='cd docs')\npyautogui.press('enter')", |
| "current_image_rel": "images/step03_traj0_r_1_2_0_current_s000164.png", |
| "next_image_rel": "images/step03_traj0_r_1_2_0_s03_neg_cd_docs_s000166.png", |
| "pixel_diff_nonzero": 5641, |
| "pixel_diff_ratio": 0.0027203896604938273 |
| } |
| }</pre></details> |
| </article> |
| <article class="card "> |
| <div><span class="badge">NEGATIVE</span> <span class="badge valid">valid diff</span></div> |
| <div class="meta"><span>wrong_navigation</span><span>diff 0.23%</span><span>s03_neg_cd_src</span></div> |
| <img src="assets/step03_traj0_r_1_2_0_s03_neg_cd_src_s000167.png" alt="next state s03_neg_cd_src"> |
| <div class="cmd">pyautogui.write(message='cd src') |
| pyautogui.press('enter')</div> |
| <details><summary>OpenCUA row JSON</summary><pre class="json">{ |
| "images": [ |
| "/NHNHOME/data/sanghyeok/qwen-cua/data/tac_gui_traj_branch_v1/images/step03_traj0_r_1_2_0_current_s000164.png" |
| ], |
| "wm_next_image": "/NHNHOME/data/sanghyeok/qwen-cua/data/tac_gui_traj_branch_v1/images/step03_traj0_r_1_2_0_s03_neg_cd_src_s000167.png", |
| "messages": [ |
| { |
| "role": "system", |
| "content": "You are a GUI agent operating a desktop. For each step, given the user task and the current screenshot, respond in this exact format:\n\n## Thought:\n<brief reasoning about progress and the chosen next local action>\n\n## Action:\n<one-sentence description of what to do>\n\n## Code:\n<pyautogui code; coordinates normalized to [0,1] when clicking>" |
| }, |
| { |
| "role": "user", |
| "content": "<image>\nWhat is the next action?" |
| }, |
| { |
| "role": "assistant", |
| "content": "## Thought:\nProgress: Inspect the requested guide file. The next local terminal action should be 'cat guide.txt'.\n\n## Action:\nType the command 'cat guide.txt' and press Enter.\n\n## Code:\npyautogui.write(message='cat guide.txt')\npyautogui.press('enter')" |
| }, |
| { |
| "role": "user", |
| "content": "<image>\nWhat is the next action?" |
| }, |
| { |
| "role": "assistant", |
| "content": "## Thought:\nProgress: Return from docs to the root folder. The next local terminal action should be 'cd ..'.\n\n## Action:\nType the command 'cd ..' and press Enter.\n\n## Code:\npyautogui.write(message='cd ..')\npyautogui.press('enter')" |
| }, |
| { |
| "role": "user", |
| "content": "Task: In the Linux terminal, inspect docs/guide.txt, return to the root folder, then enter the data folder.\n<image>\nWhat is the next action?" |
| }, |
| { |
| "role": "assistant", |
| "content": "## Thought:\nThis is a plausible but incorrect local action for the same state: 'cd src'.\n\n## Action:\nType the command 'cd src' and press Enter.\n\n## Code:\npyautogui.write(message='cd src')\npyautogui.press('enter')" |
| } |
| ], |
| "meta": { |
| "source": "tac_gui_traj_branch_v1", |
| "base_source": "cf_terminal_traj_fhd", |
| "trajectory_id": "tac_terminal_docs_data_v1", |
| "state_key": "tac_terminal_docs_data_v1/step3", |
| "base_state_key": "traj0/r.1.2.0", |
| "step_index": 3, |
| "task": "In the Linux terminal, inspect docs/guide.txt, return to the root folder, then enter the data folder.", |
| "step_intent": "Enter the data folder after returning to root.", |
| "branch_id": "s03_neg_cd_src", |
| "branch_idx": 2, |
| "branch_type": "wrong_navigation", |
| "is_gold": false, |
| "gold_command": "cd data", |
| "command": "cd src", |
| "action_granularity": "policy_step_local_action", |
| "code_norm": "pyautogui.write(message='cd src')\npyautogui.press('enter')", |
| "current_image_rel": "images/step03_traj0_r_1_2_0_current_s000164.png", |
| "next_image_rel": "images/step03_traj0_r_1_2_0_s03_neg_cd_src_s000167.png", |
| "pixel_diff_nonzero": 4864, |
| "pixel_diff_ratio": 0.002345679012345679 |
| } |
| }</pre></details> |
| </article> |
| <article class="card "> |
| <div><span class="badge">NEGATIVE</span> <span class="badge valid">valid diff</span></div> |
| <div class="meta"><span>wrong_file_read</span><span>diff 0.18%</span><span>s03_neg_cat_notes_md</span></div> |
| <img src="assets/step03_traj0_r_1_2_0_s03_neg_cat_notes_md_s000168.png" alt="next state s03_neg_cat_notes_md"> |
| <div class="cmd">pyautogui.write(message='cat notes.md') |
| pyautogui.press('enter')</div> |
| <details><summary>OpenCUA row JSON</summary><pre class="json">{ |
| "images": [ |
| "/NHNHOME/data/sanghyeok/qwen-cua/data/tac_gui_traj_branch_v1/images/step03_traj0_r_1_2_0_current_s000164.png" |
| ], |
| "wm_next_image": "/NHNHOME/data/sanghyeok/qwen-cua/data/tac_gui_traj_branch_v1/images/step03_traj0_r_1_2_0_s03_neg_cat_notes_md_s000168.png", |
| "messages": [ |
| { |
| "role": "system", |
| "content": "You are a GUI agent operating a desktop. For each step, given the user task and the current screenshot, respond in this exact format:\n\n## Thought:\n<brief reasoning about progress and the chosen next local action>\n\n## Action:\n<one-sentence description of what to do>\n\n## Code:\n<pyautogui code; coordinates normalized to [0,1] when clicking>" |
| }, |
| { |
| "role": "user", |
| "content": "<image>\nWhat is the next action?" |
| }, |
| { |
| "role": "assistant", |
| "content": "## Thought:\nProgress: Inspect the requested guide file. The next local terminal action should be 'cat guide.txt'.\n\n## Action:\nType the command 'cat guide.txt' and press Enter.\n\n## Code:\npyautogui.write(message='cat guide.txt')\npyautogui.press('enter')" |
| }, |
| { |
| "role": "user", |
| "content": "<image>\nWhat is the next action?" |
| }, |
| { |
| "role": "assistant", |
| "content": "## Thought:\nProgress: Return from docs to the root folder. The next local terminal action should be 'cd ..'.\n\n## Action:\nType the command 'cd ..' and press Enter.\n\n## Code:\npyautogui.write(message='cd ..')\npyautogui.press('enter')" |
| }, |
| { |
| "role": "user", |
| "content": "Task: In the Linux terminal, inspect docs/guide.txt, return to the root folder, then enter the data folder.\n<image>\nWhat is the next action?" |
| }, |
| { |
| "role": "assistant", |
| "content": "## Thought:\nThis is a plausible but incorrect local action for the same state: 'cat notes.md'.\n\n## Action:\nType the command 'cat notes.md' and press Enter.\n\n## Code:\npyautogui.write(message='cat notes.md')\npyautogui.press('enter')" |
| } |
| ], |
| "meta": { |
| "source": "tac_gui_traj_branch_v1", |
| "base_source": "cf_terminal_traj_fhd", |
| "trajectory_id": "tac_terminal_docs_data_v1", |
| "state_key": "tac_terminal_docs_data_v1/step3", |
| "base_state_key": "traj0/r.1.2.0", |
| "step_index": 3, |
| "task": "In the Linux terminal, inspect docs/guide.txt, return to the root folder, then enter the data folder.", |
| "step_intent": "Enter the data folder after returning to root.", |
| "branch_id": "s03_neg_cat_notes_md", |
| "branch_idx": 3, |
| "branch_type": "wrong_file_read", |
| "is_gold": false, |
| "gold_command": "cd data", |
| "command": "cat notes.md", |
| "action_granularity": "policy_step_local_action", |
| "code_norm": "pyautogui.write(message='cat notes.md')\npyautogui.press('enter')", |
| "current_image_rel": "images/step03_traj0_r_1_2_0_current_s000164.png", |
| "next_image_rel": "images/step03_traj0_r_1_2_0_s03_neg_cat_notes_md_s000168.png", |
| "pixel_diff_nonzero": 3726, |
| "pixel_diff_ratio": 0.001796875 |
| } |
| }</pre></details> |
| </article></div></div> |
| </div> |
| </section> |
| </section> |
| </main> |
| </body></html> |
