File size: 17,020 Bytes
cfcbbc8 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 |
import os
import pandas as pd
import re
import glob
from tqdm import tqdm
import datetime
import openai
import argparse
import io
def summarize_results(results_dirs, output_csv, model, no_llm = False):
client = openai.OpenAI(
api_key = os.environ.get('CBORG_API_KEY'),
base_url = 'https://api.cborg.lbl.gov'
)
error_description_prompt = (
"You are an expert assistant. Below is a comprehensive log of a multi-step workflow from a high energy physics analysis framework.\n\n"
"The workflow includes:\n"
"- A user provides an analysis task prompt.\n"
"- A supervisor agent breaks down the task and instructs a coder agent.\n"
"- The coder agent generates code, which is executed.\n"
"- The supervisor reviews results and may iterate with the coder to fix issues until the task is complete.\n"
"The log contains the user prompt, supervisor/coder dialogue, code, and execution outputs for all iterations.\n\n"
"Your task: Summarize all errors encountered during the entire workflow in clear, concise language. "
"Do NOT repeat or quote the log, prompt, or instructions. "
"Do NOT include code, explanations, or any text except your error summary.\n\n"
"For each error, use the following structure:\n"
"- Error Type: [brief description of the nature of the error]\n"
"- Cause: [if identifiable]\n"
"- Responsible Party: [user, supervisor, coder, or external]\n"
"- Consequence: [result or impact]\n"
"- Context: [any important context]\n"
"- Workflow Response: [Did the supervisor diagnose and address it?"
"Did the coder attempt a fix? Was the fix successful, unsuccessful, or misdiagnosed?"
"Was the error ignored or did it persist? Summarize the recovery process and its outcome for each error.]\n"
"List each error as a separate bullet point using this template.\n"
"If there is a validation error, look in the validation log and use the same structure to identify the causes of the validation error."
"If no errors occurred, respond: 'No errors found.'\n"
"Do NOT include code, explanations, or any text except your error summary.\n"
"Limit your entire summary to 3000 characters. "
"If no errors occurred, respond: 'No errors found.'\n\n"
)
results = []
for results_dir in results_dirs:
for name in tqdm(os.listdir(results_dir), desc=f"generating error descriptions for {results_dir}"):
output_dir = os.path.join(results_dir, name)
if os.path.isdir(output_dir):
# Extract config (everything before "_step")
config_match = re.match(r'^(.*?)_step\d+', name)
config = config_match.group(1) if config_match else None
# Extract step (int after "_step")
step_match = re.search(r'_step(\d+)', name)
step = int(step_match.group(1)) if step_match else None
result = {
"supervisor": None,
"coder": None,
"step": step,
"success": None,
"iterations": None,
"duration": None,
"API_calls": None,
"input_tokens": None,
"output_tokens": None,
"user_prompt_tokens": None,
"supervisor_to_coder_tokens": None,
"coder_output_tokens": None,
"feedback_to_supervisor_tokens": None,
"error": "Uncategorized",
"error_description": None,
"output_dir": output_dir,
}
log_dir = os.path.join(output_dir, "logs")
if os.path.isdir(log_dir):
comp_log_files = glob.glob(os.path.join(log_dir, "*comprehensive_log.txt"))
comp_log_str = None
if comp_log_files:
with open(comp_log_files[0], "r") as f:
comp_log_str = f.read()
else:
result["success"] = False
result["error_description"] = "comprehensive log file not found"
results.append(result)
continue
supervisor_match = re.search(r"Supervisor:\s*([^\s]+)", comp_log_str)
coder_match = re.search(r"Coder:\s*([^\s]+)", comp_log_str)
if supervisor_match:
result["supervisor"] = supervisor_match.group(1)
if coder_match:
result["coder"] = coder_match.group(1)
iterations_match = re.search(r"Total Iterations:\s*(\d+)", comp_log_str)
if iterations_match:
result["iterations"] = int(iterations_match.group(1))
duration_match = re.search(r"Duration:\s*([0-9:.\s]+)", comp_log_str)
if duration_match:
duration_str = duration_match.group(1).strip()
try:
t = datetime.datetime.strptime(duration_str, "%H:%M:%S.%f")
except ValueError:
t = datetime.datetime.strptime(duration_str, "%H:%M:%S")
result["duration"] = t.hour * 3600 + t.minute * 60 + t.second + t.microsecond / 1e6
api_calls_match = re.search(r"Total API Calls:\s*(\d+)", comp_log_str)
if api_calls_match:
result["API_calls"] = int(api_calls_match.group(1))
input_tokens_match = re.search(r"Total Input Tokens:\s*(\d+)", comp_log_str)
if input_tokens_match:
result["input_tokens"] = int(input_tokens_match.group(1))
output_tokens_match = re.search(r"Total Output Tokens:\s*(\d+)", comp_log_str)
if output_tokens_match:
result["output_tokens"] = int(output_tokens_match.group(1))
match = re.search(r"User Prompt Tokens:\s*(\d+)", comp_log_str)
if match:
result["user_prompt_tokens"] = int(match.group(1))
match = re.search(r"Supervisor to Coder Tokens:\s*(\d+)", comp_log_str)
if match:
result["supervisor_to_coder_tokens"] = int(match.group(1))
match = re.search(r"Coder Output Tokens:\s*(\d+)", comp_log_str)
if match:
result["coder_output_tokens"] = int(match.group(1))
match = re.search(r"Feedback to Supervisor Tokens:\s*(\d+)", comp_log_str)
if match:
result["feedback_to_supervisor_tokens"] = int(match.group(1))
# Check validation.log to see if outputs are correct
val_log_files = glob.glob(os.path.join(log_dir, "*validation.log"))
val_log_str = None
if val_log_files:
with open(val_log_files[0], "r") as f:
val_log_str = f.read()
matches = re.findall(r'(✅ Validation successful|❌ Validation failed)', val_log_str)
if not matches:
result["success"] = False
else:
last = matches[-1]
result["success"] = last == "✅ Validation successful"
if (no_llm):
if (result["success"]):
result["error"] = None
else:
result["error"] = "Validation Error"
val_log_str = val_log_str.replace('\n', '').replace('\r', '')
else:
result["success"] = False
val_log_str = ""
if (not no_llm):
try:
response = client.chat.completions.create(
model = model,
messages = [
{
'role': 'user',
'content': error_description_prompt +
"\nComprehensive Log:\n" + comp_log_str +
"\nValidation Log:\n" + val_log_str
}
],
temperature = 0.0
)
error_description = response.choices[-1].message.content
error_description = " ".join(error_description.split())
error_description = error_description[:3000]
result["error_description"] = error_description
except Exception as e:
print(f"OpenAI API error: {e}")
else:
if ("API call failed" in comp_log_str):
result["error"] = "API Call Error"
else:
result["success"] = False
result["error_description"] = "job submission failure"
results.append(result)
df = pd.DataFrame(results)
df = df.sort_values(by=["supervisor", "coder", "step", "output_dir"])
df.to_csv(output_csv, index=False)
print(f"Results written to {output_csv}")
def categorize_errors(output_csv, model):
client = openai.OpenAI(
api_key = os.environ.get('CBORG_API_KEY'),
base_url = 'https://api.cborg.lbl.gov'
)
# Load the CSV as a pandas DataFrame
df = pd.read_csv(output_csv, comment='#')
# Get list of error_descriptions and their indices (for mapping back)
error_descriptions = df['error_description'].fillna("").tolist()
# 1. Generate categories prompt
create_categories_prompt = (
"You are an expert at analyzing and organizing error messages from machine learning workflows in high energy physics.\n\n"
"Workflow summary:\n"
"- A user provides an analysis task prompt.\n"
"- A supervisor agent breaks down the task and instructs a coder agent.\n"
"- The coder agent generates code, which is executed.\n"
"- The supervisor reviews results and may iterate with the coder to fix issues until the task is complete.\n"
"Error descriptions below are collected from all steps and iterations of this workflow.\n\n"
"Your task: Identify 5 to 10 distinct, meaningful categories that best capture the underlying nature or root cause of the errors in the list. "
"Focus on grouping errors by what fundamentally caused them (such as logic mistakes, miscommunication, missing dependencies, data mismatches, etc.), "
"rather than by their symptoms, error messages, or observable effects. "
"Do NOT create categories based on how the error was observed or reported, but on the underlying issue that led to it.\n\n"
"Each category should have a short, clear name and a one-sentence description that explains what kinds of errors belong in that category.\n\n"
"Output only the categories in this format:\n"
"1. [Category Name]: [One-sentence description]\n"
"2. [Category Name]: [One-sentence description]\n"
"...\n"
"N. [Category Name]: [One-sentence description]\n\n"
"Here are some example error categories:\n"
"- Coding API Error: the coder incorrectly utilized common python packages (e.g. numpy, awkward, uproot, pandas)\n"
"- User Prompt Misunderstanding: the supervisor did not properly interpret the user prompt"
"Here are some error descriptions after running the workflow:\n"
"```\n"
)
# Add error descriptions to prompt, one per line
create_categories_prompt += "\n".join(error_descriptions) + "\n```"
# 2. Call LLM to get categories
try:
response = client.chat.completions.create(
model=model,
messages=[{'role': 'user', 'content': create_categories_prompt}],
temperature=0.0
)
error_categories = response.choices[-1].message.content.strip()
print("Categories found by LLM:\n", error_categories)
except Exception as e:
print(f"LLM API error (category generation): {e}")
return
df['error'] = df['error'].astype(str)
for idx, error_description in tqdm(enumerate(error_descriptions), total=len(error_descriptions), desc="categorizing errors"):
if not error_description.strip():
continue
categorize_errors_prompt = (
"You are an expert at classifying error messages from machine learning workflows in high energy physics.\n\n"
"Workflow summary:\n"
"- A user provides an analysis task prompt.\n"
"- A supervisor agent breaks down the task and instructs a coder agent.\n"
"- The coder agent generates code, which is executed.\n"
"- The supervisor reviews results and may iterate with the coder to fix issues until the task is complete.\n"
"The error descriptions below are collected from all steps and iterations of this workflow.\n\n"
"Below is a list of error categories, each with a short description:\n"
f"{error_categories}\n\n"
"Your task: For the given error description, select the single most appropriate error category from the list above. "
"Base your choice on the underlying nature or root cause of the error, not on the symptoms, error messages, or observable effects. "
"Focus on what fundamentally caused the error, such as logic mistakes, missing dependencies, data mismatches, or miscommunication, rather than how the error was reported or observed.\n"
"Return ALL applicable category names, each wrapped with three asterisks on each side, separated by commas, like this: ***Category One***, ***Category Two***"
"Do not include any other text, explanation, or formatting."
"Error description:\n"
"```\n"
f"{error_description}\n"
"```"
)
def parse_categories(llm_output):
# Find all ***Category Name*** matches
return [cat.strip() for cat in re.findall(r"\*\*\*(.*?)\*\*\*", llm_output)]
try:
response = client.chat.completions.create(
model=model,
messages=[{'role': 'user', 'content': categorize_errors_prompt}],
temperature=0.0
)
assignments_text = response.choices[-1].message.content.strip()
categories = parse_categories(assignments_text)
df.at[idx, 'error_categories'] = categories if categories else ["Uncategorized"]
except Exception as e:
print(f"LLM API error (assignment) at row {idx}: {e}")
df.at[idx, 'error'] = "LLM API error"
df.to_csv(output_csv, index=False)
with open(output_csv, 'w', encoding='utf-8') as f:
f.write("# LLM Generated Error Categories:\n")
for line in error_categories.splitlines():
f.write(f"# {line}\n")
f.write("\n")
df.to_csv(f, index=False)
print(f"Saved categorized errors to {output_csv}")
def main():
parser = argparse.ArgumentParser(description="Summarize experiment logs and errors")
parser.add_argument("--results_dir", type=str, default=" ", nargs='+', required=False, help="One or more directories containing experiment results")
parser.add_argument("--output_csv", type=str, default="results_summary.csv", help="Path to output CSV file")
parser.add_argument("--model", type=str, default="gpt-oss-120b", help="LLM model to use for error summarization")
parser.add_argument("--no_llm", action="store_true", default=False, help="If set, only generate the CSV without LLM error description or categorization")
args = parser.parse_args()
summarize_results(
results_dirs=args.results_dir,
output_csv=args.output_csv,
model=args.model,
no_llm=args.no_llm
)
if not args.no_llm:
categorize_errors(
output_csv=args.output_csv,
model=args.model
)
else:
print("LLM error description and categorization skipped (--no_llm set)")
if __name__ == "__main__":
main() |