This commit is contained in:
yuyr 2025-06-17 14:23:59 +08:00
parent f34e4a464d
commit a5b060d4f8
2 changed files with 187 additions and 1 deletions

View File

@ -7,7 +7,10 @@
# 工具
- `scripts`
* portforward.sh 将电商网站的mysql端口映射到本机
* portforward.sh 将电商网站的mysql端口映射到本机, 测试连接 mysql -u root -P 23306 -p 1234567890
# conda幻觉
- qa311
# 当前进展:
- 对电商后台网站使用方法c构造1k条问答对数据集cms1k转换成WebArena任务描述规范可以对WebRL 模型进行rollout评测

View File

@ -0,0 +1,183 @@
"""
分析多轮实验的结果中eige renwu 每个任务的平均成功率
每轮结果分别在不同的目录下resutls/webrlvr_cms1k_easy_0, results/webrlvr_cms1k_easy_1, ...
每个目录下有一个actions目录每个actions目录下有多个json文件, 0.json, 1.json等 数字为任务id每个json文件中记录了每个任务的执行结果
0.json文件内容示例
{
"task_id": 0,
"score": 0.0,
"actions": [
"# Element: the 'Bestsellers' tab in the left sidebar menu\ndo(action=\"Click\", element=\"21\")",
"# Element: the 'Most Viewed Products' link in the left sidebar under the 'Reports' section\ndo(action=\"Click\", element=\"22\")",
"exit(message=\"N/A\")"
]
}
score 为0.0表示失败为1.0表示成功
总共1003个任务, 从0到1002
总共有4轮实验分别是0, 1, 2, 3
现在要分别统计每个任务在所有轮次的成功次数失败次数成功率
结果保存到一个json文件中
"""
import json
import os
import logging
# Configure logging
logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')
def load_task_descriptions():
"""Load task descriptions from the config file."""
config_path = "../VAB-WebArena-Lite/config_files/wa/webrlvr_cms1k_easy.json"
try:
with open(config_path, 'r') as f:
tasks = json.load(f)
return {str(task['task_id']): task['intent'] for task in tasks}
except Exception as e:
logging.error(f"Error loading task descriptions: {e}")
return {}
def analyze_results():
"""
Analyzes multi-round experiment results to calculate average success rates for each task.
"""
logging.info("Starting analysis...")
num_rounds = 4
num_tasks = 1003
base_path = "results"
experiment_name = "webrlvr_cms1k_easy"
output_filename = "analysis_results.json"
stats_output_filename = "success_group_stats.json"
# Load task descriptions
task_descriptions = load_task_descriptions()
task_stats = {
str(i): {"success": 0, "failure": 0} for i in range(num_tasks)
}
for i in range(num_rounds):
round_dir = os.path.join(base_path, f"{experiment_name}_{i}", "actions")
if not os.path.exists(round_dir):
logging.warning(f"Directory not found, skipping: {round_dir}")
continue
logging.info(f"Processing round {i} from {round_dir}")
for task_id in range(num_tasks):
file_path = os.path.join(round_dir, f"{task_id}.json")
task_id_str = str(task_id)
if not os.path.exists(file_path):
task_stats[task_id_str]["failure"] += 1
continue
try:
with open(file_path, "r") as f:
data = json.load(f)
if data.get("score") == 1.0:
task_stats[task_id_str]["success"] += 1
else:
task_stats[task_id_str]["failure"] += 1
except (json.JSONDecodeError, KeyError) as e:
logging.error(f"Error reading or parsing {file_path}: {e}")
task_stats[task_id_str]["failure"] += 1
logging.info("Calculating success rates...")
for task_id, stats in task_stats.items():
total_runs = stats["success"] + stats["failure"]
if total_runs > 0:
stats["success_rate"] = stats["success"] / total_runs
else:
stats["success_rate"] = 0.0
logging.info("Grouping tasks by success count...")
success_groups = {i: [] for i in range(num_rounds + 1)}
# Create dictionaries to track successful passes and their trace lengths for each task
task_successful_passes = {str(i): [] for i in range(num_tasks)}
task_successful_trace_lengths = {str(i): [] for i in range(num_tasks)}
# First pass: collect successful passes and trace lengths for each task
for i in range(num_rounds):
round_dir = os.path.join(base_path, f"{experiment_name}_{i}", "actions")
if not os.path.exists(round_dir):
continue
for task_id in range(num_tasks):
file_path = os.path.join(round_dir, f"{task_id}.json")
task_id_str = str(task_id)
if os.path.exists(file_path):
try:
with open(file_path, "r") as f:
data = json.load(f)
if data.get("score") == 1.0:
task_successful_passes[task_id_str].append(i)
# Get the length of actions array for successful passes
trace_length = len(data.get("actions", []))
task_successful_trace_lengths[task_id_str].append(trace_length)
except (json.JSONDecodeError, KeyError):
pass
# Second pass: group tasks by success count
for task_id, stats in task_stats.items():
success_count = stats["success"]
success_groups[success_count].append(task_id)
# Create detailed success group statistics
success_group_stats = {}
for success_count, task_ids in success_groups.items():
group_info = {
"count": len(task_ids),
"task_ids": task_ids
}
# Add task descriptions for groups with success count > 0
if success_count > 0:
task_details = []
for task_id in task_ids:
task_detail = {
"task_id": task_id,
"description": task_descriptions.get(task_id, "Description not found"),
"successful_passes": task_successful_passes[task_id],
"successful_trace_lengths": task_successful_trace_lengths[task_id] # Add trace lengths
}
task_details.append(task_detail)
group_info["task_details"] = task_details
success_group_stats[success_count] = group_info
# Print detailed information for each success group
logging.info("\n=== Success Group Statistics ===")
for success_count, group_info in success_group_stats.items():
logging.info(f"\nSuccess Count {success_count}:")
logging.info(f"Number of tasks: {group_info['count']}")
logging.info(f"Task IDs: {group_info['task_ids']}")
if success_count > 0:
logging.info("Task Descriptions:")
for task in group_info['task_details']:
logging.info(f"Task {task['task_id']}: {task['description']}")
logging.info(f"Successful passes: {task['successful_passes']}")
logging.info(f"Trace lengths: {task['successful_trace_lengths']}")
# Save success group statistics to separate file
with open(stats_output_filename, "w") as f:
json.dump(success_group_stats, f, indent=4)
logging.info(f"\nSuccess group statistics saved to {stats_output_filename}")
# Save original analysis results
output_data = {
"task_stats": task_stats,
"success_groups": success_groups,
"success_group_counts": {count: len(tasks) for count, tasks in success_groups.items()},
}
with open(output_filename, "w") as f:
json.dump(output_data, f, indent=4)
logging.info(f"Analysis complete. Results saved to {output_filename}")
if __name__ == "__main__":
analyze_results()