Benchmark Results Report
+ + +
+
+
+
+
+
+
+"""
+
+ # Process each Task/Step data
+ for i, step_data in enumerate(data):
+ task_id = step_data.get("task_id", "N/A")
+ step_id = step_data.get("step_id", "N/A")
+ intent = step_data.get("intent", "N/A")
+ start_url = step_data.get("start_url", "N/A")
+ gt_checklist = step_data.get("gt_checklist", "N/A")
+ generated_checklist = step_data.get("generated_checklist", None)
+ trajectory = step_data.get("trajectory", "N/A")
+ text_observation = step_data.get("text_observation", "N/A")
+ source_name = step_data.get("source_name", "")
+
+ # Wrap each Task/Step in a container with a unique ID (hidden initially)
+ html_content += f"""
+
+
+
+
+
+
+"""
+
+ # Save the HTML file
+ try:
+ with open(html_path, 'w', encoding='utf-8') as f:
+ f.write(html_content)
+ print(f"Completed: HTML report created at {html_path}")
+ except IOError:
+ print(f"Error: Failed to write HTML file - {html_path}")
+ except Exception as e:
+ print(f"Unexpected error during HTML file saving: {e}")
+
+# --- Example Usage ---
+# input_json_file = 'path/to/your/results.json'
+# output_html_file = 'trajectory_report.html'
+# create_html_report(input_json_file, output_html_file)
\ No newline at end of file
diff --git a/agent/reward.py b/agent/reward.py
new file mode 100644
index 0000000000000000000000000000000000000000..2337bc24110a83634c0ec9d6feba3edde13b85f8
--- /dev/null
+++ b/agent/reward.py
@@ -0,0 +1,96 @@
+import time
+from typing import List, Dict, Any, Optional, Union
+import numpy as np
+from .mini_bench.reward_agent import ProgressJudgeAgent
+from .reward_postprocessor import REWARD_PROCESSORS, REWARD_PROCESSOR_N_SAMPLES, extract_judge_hash
+import json
+import os
+from concurrent.futures import ThreadPoolExecutor, as_completed
+
+def _process_unit(idx, unit, configs, n_samples, reward_processor, max_retries=5):
+ """하나의 unit을 처리해 (idx, reward, thought)를 돌려준다."""
+ agent = ProgressJudgeAgent(configs)
+ current_temperature = configs["temperature"]
+
+ rewards = []
+ n_err = 0
+ retry_count = 0
+ judge_hash_count_thought = {}
+
+ while len(rewards) < n_samples and retry_count < max_retries:
+ # 외부 API 호출
+ responses, _ = agent.generate_probs(
+ unit, "ours", n=n_samples - len(rewards), temperature=current_temperature
+ )
+
+ for response in responses:
+ content = response["response"]
+ thought = content # 전체를 로그로 저장
+ reward = REWARD_PROCESSORS[reward_processor](response)
+ rewards.append(reward)
+
+ if np.isnan(reward) or reward is None:
+ n_err += 1
+ else:
+ judge_hash = extract_judge_hash(response)
+ judge_hash_count_thought[judge_hash] = (judge_hash_count_thought.get(judge_hash, (0, None))[0] + 1, thought)
+
+ if n_err > 0:
+ # 실패 시 온도를 높여 재시도
+ if n_samples == 1:
+ current_temperature = 0.5
+ retry_count += 1
+
+ reward = np.nanmean(rewards)
+ if np.isnan(reward):
+ print(f"[idx={idx}] Warning: reward is NaN after retries -> set 0")
+ reward = 0.0
+ print(judge_hash_count_thought)
+ thought = max(judge_hash_count_thought.values(), key=lambda x: x[0])[1]
+
+ return idx, reward, thought
+
+
+def get_ar_reward(dataset, base_url, model_name, reward_processor='avg_logits', max_workers=8):
+ """원본 get_ar_reward를 스레드 버전으로 교체."""
+ n_samples = REWARD_PROCESSOR_N_SAMPLES[reward_processor]
+
+ temperature = 0.5 if n_samples > 1 else 0.0
+
+ configs = {
+ "model_name": model_name,
+ "base_url": base_url,
+ "api_key": "empty",
+ "temperature": temperature,
+ "num_generate": 1,
+ "use_checklist": True,
+ "input_type": "text_only",
+ "text_obs_type": "axtree",
+ "image_obs_type": "som",
+ "use_in_progress": True,
+ "use_multimodal": False,
+ "use_log_probs": True,
+ }
+
+ t_start = time.time()
+ results = [None] * len(dataset)
+
+ with ThreadPoolExecutor(max_workers=max_workers) as executor:
+ futures = [
+ executor.submit(
+ _process_unit, idx, unit, configs, n_samples, reward_processor
+ )
+ for idx, unit in enumerate(dataset)
+ ]
+
+ for fut in as_completed(futures):
+ idx, reward, thought = fut.result()
+ results[idx] = (reward, thought)
+
+ # 순서 보존된 리스트로 분리
+ final_rewards = [float(r) for r, _ in results]
+ thoughts = [t for _, t in results]
+
+ print(f"Time taken (threaded): {time.time() - t_start:.2f} s")
+ return final_rewards, thoughts
+
diff --git a/agent/reward_postprocessor.py b/agent/reward_postprocessor.py
new file mode 100644
index 0000000000000000000000000000000000000000..4124dba8cf12745a64d756002c3c5e99a17f1123
--- /dev/null
+++ b/agent/reward_postprocessor.py
@@ -0,0 +1,41 @@
+import numpy as np
+import re
+
+
+def extract_judge_hash(response):
+ """
+ checklist 별로 yes, in, no를 판단한 정보를 hash 형태로 변환하여 반환
+ """
+ content = response['response']
+
+ try:
+ judge_content = content.lower().replace(' ', '').split('
+
+"""
+
+ # Finalize HTML and add JavaScript
+ html_content += """
+
+
+""" + + # Chosen Responses + if 'chosen' in step_data and step_data['chosen']: + html_content += '' # End chosen-section
+
+ # Rejected Responses
+ if 'rejected' in step_data and step_data['rejected']:
+ html_content += '' # End rejected-section
+
+ html_content += """
+
+ Task ID: {html.escape(str(task_id))} | Step ID: {html.escape(str(step_id))} {f'({html.escape(source_name)})' if source_name else ''}
+Intent:
+{html.escape(intent)}
+Start URL: {html.escape(start_url)}
+ +Ground Truth Checklist:
+{html.escape(gt_checklist)}
+"""
+ if checklist_generation and generated_checklist is not None:
+ html_content += f"""
+
+
+"""
+
+ html_content += f"""
+ Generated Checklist (Click to expand/collapse)
+{html.escape(str(generated_checklist))}
+
+
+
+ Trajectory (Click to expand/collapse)
+{html.escape(trajectory)}
+
+
+ Text Observation (Click to expand/collapse)
+{html.escape(text_observation)}
+ +""" + + # Chosen Responses + if 'chosen' in step_data and step_data['chosen']: + html_content += '
Chosen Responses:
' + for choice_block in step_data['chosen']: + thought = choice_block.get('thought', 'N/A') + action = choice_block.get('action', 'N/A') + responses = choice_block.get('response', []) + scores = choice_block.get('score', []) + + # Add Thought and Action information + html_content += f""" +
+
"""
+
+ # Loop through responses and create toggles
+ for idx, (response, score) in enumerate(zip(responses, scores)):
+ html_content += f"""
+ Thought:
+{html.escape(thought)}
+ Action:
+{html.escape(action)}
+
+
"""
+ html_content += 'Judge Response {idx + 1}: {html.escape(str(score))}
+{html.escape(str(response))}
+ Rejected Responses:
' + for rejection_block in step_data['rejected']: + thought = rejection_block.get('thought', 'N/A') + action = rejection_block.get('action', 'N/A') + responses = rejection_block.get('response', []) + scores = rejection_block.get('score', []) + + # Add Thought and Action information + html_content += f""" +
+
"""
+
+ # Loop through responses and create toggles
+ for idx, (response, score) in enumerate(zip(responses, scores)):
+ html_content += f"""
+ Thought:
+{html.escape(thought)}
+ Action:
+{html.escape(action)}
+
+
"""
+ html_content += 'Judge Response {idx + 1}: {html.escape(str(score))}
+{html.escape(str(response))}
+
+ """
+
+ for idx, card in enumerate(rm_cards):
+ rm_cards_html += CARD_HTML_TEMPLATE.format(
+ additional_class='top-candidate' if idx == 0 else '',
+ k=idx+1,
+ suffix='(best)' if idx == 0 else '',
+ thought=card['thought'],
+ action=card['action'],
+ reward=card['reward'],
+ feedback=card['feedback']
+ )
+
+ rm_cards_html += "
"
+ all_trajectory = trajectory
+
+ # Create HTML for trajectory display
+ trajectory_html = f"""
+
+
+ """
+
+ for idx, (after_img, cands) in enumerate(all_trajectory):
+ # Convert image to base64 if needed
+ img = all_trajectory[idx][0]
+ if isinstance(img, np.ndarray):
+ img = Image.fromarray(img)
+ if isinstance(img, Image.Image):
+ buffer = io.BytesIO()
+ img.save(buffer, format="JPEG")
+ img_str = base64.b64encode(buffer.getvalue()).decode()
+ img_src = f"data:image/jpeg;base64,{img_str}"
+ else:
+ img_src = img
+
+ trajectory_html += f"""
+
"
+
+ last_checklist_view, last_trajectory_html = checklist_view, trajectory_html
+ yield state, screenshot, last_checklist_view, rm_cards_html, last_trajectory_html
+ yield state, screenshot, last_checklist_view, rm_cards_html, last_trajectory_html
diff --git a/requirements.txt b/requirements.txt
index 5582dc4fb19a373861662822e075bcd4cf6ff9a0..bb8e47a583bb7b74c8504053dbb375c759cbc3ea 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -1,4 +1,7 @@
gradio
openai
numpy
-gymnasium
\ No newline at end of file
+gymnasium
+langsmith
+langchain[openai]
+pillow
\ No newline at end of file
diff --git a/templates/card.html b/templates/card.html
new file mode 100644
index 0000000000000000000000000000000000000000..274a06f43b3b615e0af8f310946c603bbc44a0dd
--- /dev/null
+++ b/templates/card.html
@@ -0,0 +1,33 @@
+
+
+ """
+
+ trajectory_html += "Step {idx + 1}
+
+
+
+ 
+
+
+
+ Action Candidates:
+
+ """
+
+ # Display all candidates for this step
+ for i, cand in enumerate(cands):
+ action = cand['action']
+ reward = cand['reward']
+
+ trajectory_html += f"""
+
+
+
+ """
+
+ trajectory_html += """
+
+ Action {i+1}{' (Selected)' if i == 0 else ''}
+ Reward: {reward}
+
+ {action}
+
+
\ No newline at end of file
diff --git a/templates/rm_cards.css b/templates/rm_cards.css
new file mode 100644
index 0000000000000000000000000000000000000000..e09b97dd6a01ae909031d4e476af4120da0b4d0b
--- /dev/null
+++ b/templates/rm_cards.css
@@ -0,0 +1,106 @@
+.rm-cards-container {
+ display: flex;
+ gap: 15px;
+ padding: 10px 0;
+ overflow-x: auto;
+}
+
+.rm-card {
+ min-width: 300px;
+ max-width: 400px;
+ border: 1px solid #ddd;
+ border-radius: 8px;
+ overflow: hidden;
+ box-shadow: 0 2px 5px rgba(0, 0, 0, 0.1);
+ background: rgba(255, 255, 255, 0.0);
+}
+
+.rm-card.top-candidate {
+ border: 3px solid #007bff;
+ box-shadow: 0 4px 8px rgba(0, 0, 150, 0.2);
+}
+
+.rm-card-header {
+ background: rgba(240, 240, 240, 0.3);
+ padding: 10px 15px;
+ font-weight: bold;
+ border-bottom: 1px solid #ddd;
+ display: flex;
+ justify-content: space-between;
+ align-items: center;
+}
+
+.top-candidate .rm-card-header {
+ background: rgba(230, 242, 255, 0.3);
+}
+
+.reward-badge {
+ background: #007bff;
+ color: white;
+ padding: 3px 8px;
+ border-radius: 12px;
+ font-size: 0.9em;
+}
+
+.rm-card-body {
+ padding: 15px;
+ background-color: rgba(240, 240, 240, 0.0);
+}
+
+.card-section {
+ margin-bottom: 12px;
+}
+
+.card-section-title {
+ font-weight: bold;
+ margin-bottom: 5px;
+ color: #555;
+}
+
+.thought-content {
+ background: rgba(247, 247, 255, 0.3);
+ border: 1px solid #d0d0ff;
+ border-radius: 6px;
+ padding: 8px;
+}
+
+.action-content {
+ background: rgba(240, 255, 240, 0.3);
+ border: 1px solid #d0ffd0;
+ border-radius: 6px;
+ padding: 8px;
+}
+
+.feedback-content {
+ background: rgba(255, 247, 240, 0.3);
+ border: 1px solid #ffd0a0;
+ border-radius: 6px;
+ padding: 8px;
+}
+
+.reward-content {
+ background: rgba(240, 240, 255, 0.3);
+ border: 1px solid #d0d0ff;
+ border-radius: 6px;
+ padding: 8px;
+}
+
+details {
+ margin-top: 5px;
+}
+
+summary {
+ cursor: pointer;
+ font-weight: bold;
+ color: #555;
+}
+
+summary:hover {
+ color: #007bff;
+}
+
+pre {
+ margin: 0;
+ white-space: pre-wrap;
+ word-break: break-word;
+}
\ No newline at end of file
diff --git a/templates/trajectory.css b/templates/trajectory.css
new file mode 100644
index 0000000000000000000000000000000000000000..5795985190beaec7a73168113106e8094a16c0cd
--- /dev/null
+++ b/templates/trajectory.css
@@ -0,0 +1,82 @@
+/* templates/trajectory.css */
+
+.trajectory-container {
+ display: flex;
+ flex-direction: column;
+ gap: 20px;
+}
+
+.step-container {
+ border: 1px solid #ddd;
+ border-radius: 8px;
+ overflow: hidden;
+ box-shadow: 0 2px 5px rgba(0, 0, 0, 0.1);
+}
+
+.step-header {
+ background: rgba(240, 240, 240, 0.0);
+ padding: 10px 15px;
+ font-weight: bold;
+ border-bottom: 1px solid #ddd;
+}
+
+.step-content {
+ display: flex;
+ gap: 15px;
+ flex-wrap: wrap;
+}
+
+.step-image {
+ flex: 1 1 40%;
+ padding: 10px;
+}
+
+.step-image img {
+ width: 100%;
+ border: 1px solid #eee;
+ border-radius: 4px;
+}
+
+.step-info {
+ flex: 1 1 55%;
+ display: flex;
+ flex-direction: column;
+ padding: 10px;
+}
+
+.action-candidates {
+ display: flex;
+ flex-direction: column;
+ gap: 10px;
+}
+
+.candidate-box {
+ background: rgba(245, 240, 255, 0.0);
+ border: 1px solid #d0d0ff;
+ border-radius: 6px;
+ padding: 10px;
+}
+
+.candidate-box.selected {
+ border: 2px solid #7030a0;
+ box-shadow: 0 2px 4px rgba(112, 48, 160, 0.1);
+}
+
+.box-title {
+ font-weight: bold;
+ margin-bottom: 5px;
+}
+
+.reward-text {
+ display: inline-block;
+ color: #555;
+ font-size: 0.9em;
+ margin-left: 8px;
+ font-style: italic;
+}
+
+pre {
+ margin: 0;
+ white-space: pre-wrap;
+ word-break: break-word;
+}
\ No newline at end of file
diff --git a/utils.py b/utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..c133195e82b192db4b825a1938ba2a1918a1457c
--- /dev/null
+++ b/utils.py
@@ -0,0 +1,41 @@
+import tokenize
+import io
+import base64
+import numpy as np
+from PIL import Image
+
+def remove_inline_comments_safe(code: str) -> str:
+ result = []
+ tokens = tokenize.generate_tokens(io.StringIO(code).readline)
+
+ last_line = -1
+ current_line = ''
+ for tok_type, tok_string, (srow, scol), (_, _), _ in tokens:
+ if srow != last_line:
+ if current_line:
+ result.append(current_line.rstrip())
+ current_line = ''
+ last_line = srow
+
+ if tok_type == tokenize.COMMENT:
+ # 주석 무시 (아무 것도 추가하지 않음)
+ continue
+
+ current_line += tok_string
+
+ if current_line:
+ result.append(current_line.rstrip())
+
+ return '\n'.join(result)
+
+
+def image_to_jpg_base64_url(image: Image.Image | np.ndarray) -> str:
+ """Return a base64 *JPEG* data‑URL from a PIL image or NumPy array."""
+ if isinstance(image, np.ndarray):
+ image = Image.fromarray(image)
+ if image.mode in {"RGBA", "LA"}:
+ image = image.convert("RGB")
+ with io.BytesIO() as buffer:
+ image.save(buffer, format="JPEG")
+ encoded: str = base64.b64encode(buffer.getvalue()).decode()
+ return f"data:image/jpeg;base64,{encoded}"
+ Top {k} Action Candidate {suffix}
+
+
+
+
+
+ Thought:
+
+
+ {thought}
+
+
+ Action:
+
+
+ {action}
+
+
+ Reward:
+
+
+ {reward}
+
+
+
+
+ Feedback (click to view)
+
+
+ {feedback}
+