File size: 10,855 Bytes
d4d998a 3c01baa d4d998a 3c01baa d4d998a 3c01baa d4d998a 3c01baa d4d998a 3c01baa d4d998a 3c01baa d4d998a 3c01baa d4d998a 3c01baa d4d998a 3c01baa d4d998a 3c01baa d4d998a 3c01baa d4d998a 3c01baa d4d998a 3c01baa d4d998a 3c01baa d4d998a 3c01baa d4d998a 3c01baa d4d998a 3c01baa d4d998a 3c01baa d4d998a 3c01baa d4d998a 3c01baa d4d998a 3c01baa d4d998a 3c01baa d4d998a 3c01baa d4d998a 3c01baa d4d998a 3c01baa d4d998a 3c01baa d4d998a 3c01baa d4d998a 3c01baa d4d998a |
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 |
"""
Populate the GuardBench leaderboard from HuggingFace datasets.
"""
import json
import os
import pandas as pd
import tempfile
from typing import Dict, Tuple, List
from glob import glob
from huggingface_hub import snapshot_download, hf_hub_download, HfApi
from datasets import load_dataset
from src.display.utils import GUARDBENCH_COLUMN, DISPLAY_COLS, CATEGORIES
from src.envs import RESULTS_DATASET_ID, TOKEN, LEADERBOARD_FILE, CACHE_PATH
from src.leaderboard.processor import leaderboard_to_dataframe, load_leaderboard_data, save_leaderboard_data, process_jsonl_submission, add_entries_to_leaderboard
def get_versioned_leaderboard_file(version="v0"):
"""
Get the versioned leaderboard file path.
"""
base_name, ext = os.path.splitext(LEADERBOARD_FILE)
return f"{base_name}_{version}{ext}"
def download_leaderboard_data(version="v0") -> bool:
"""
Download the latest leaderboard data from HuggingFace.
Args:
version: The dataset version to download
"""
try:
# Create a temporary directory to download the submissions
temp_dir = os.path.join(CACHE_PATH, f"temp_submissions_{version}")
os.makedirs(temp_dir, exist_ok=True)
# Get the versioned leaderboard file
leaderboard_file = get_versioned_leaderboard_file(version)
# Download the entire repository
try:
snapshot_path = snapshot_download(
repo_id=RESULTS_DATASET_ID,
repo_type="dataset",
local_dir=temp_dir,
token=TOKEN,
ignore_patterns=["*.md", ".*"],
etag_timeout=30
)
# Process all submission files
all_entries = []
submission_files = []
# Look for submission files in the submissions directory
submissions_dir = os.path.join(snapshot_path, "submissions")
version_submissions_dir = os.path.join(snapshot_path, f"submissions_{version}")
# Check both standard and versioned submission directories
if os.path.exists(submissions_dir):
submission_files.extend(glob(os.path.join(submissions_dir, "*.jsonl")))
if os.path.exists(version_submissions_dir):
submission_files.extend(glob(os.path.join(version_submissions_dir, "*.jsonl")))
# Also look for any versioned JSONL files in the root
submission_files.extend(glob(os.path.join(snapshot_path, f"*_{version}.jsonl")))
# If we're looking for v0 and no versioned files found, use generic ones
if version == "v0" and not submission_files:
submission_files.extend(glob(os.path.join(snapshot_path, "*.jsonl")))
# Process each submission file
for file_path in submission_files:
entries, _ = process_jsonl_submission(file_path)
# Filter entries to those that match the version or don't have version specified
filtered_entries = [
entry for entry in entries
if entry.get("version", "v0") == version or "version" not in entry
]
all_entries.extend(filtered_entries)
# Create leaderboard data structure
leaderboard_data = {
"entries": all_entries,
"last_updated": pd.Timestamp.now().isoformat(),
"version": version
}
# Save to local file
save_leaderboard_data(leaderboard_data, leaderboard_file)
return True
except Exception as e:
print(f"Error downloading repository: {e}")
# If we can't download the repository, try to download individual files
try:
api = HfApi(token=TOKEN)
files = api.list_repo_files(repo_id=RESULTS_DATASET_ID, repo_type="dataset")
# Look for versioned and regular files
submission_files = [
f for f in files
if (f.endswith(f'_{version}.jsonl') or
f.startswith(f'submissions_{version}/') or
(version == "v0" and f.endswith('.jsonl')))
]
all_entries = []
for file_path in submission_files:
try:
local_path = hf_hub_download(
repo_id=RESULTS_DATASET_ID,
filename=file_path,
repo_type="dataset",
token=TOKEN
)
entries, _ = process_jsonl_submission(local_path)
# Filter entries to those that match the version or don't have version specified
filtered_entries = [
entry for entry in entries
if entry.get("version", "v0") == version or "version" not in entry
]
all_entries.extend(filtered_entries)
except Exception as file_error:
print(f"Error downloading file {file_path}: {file_error}")
# Create leaderboard data structure
leaderboard_data = {
"entries": all_entries,
"last_updated": pd.Timestamp.now().isoformat(),
"version": version
}
# Save to local file
save_leaderboard_data(leaderboard_data, leaderboard_file)
return True
except Exception as list_error:
print(f"Error listing repository files: {list_error}")
# If we can't download anything, create an empty leaderboard
if not os.path.exists(leaderboard_file):
empty_data = {
"entries": [],
"last_updated": pd.Timestamp.now().isoformat(),
"version": version
}
save_leaderboard_data(empty_data, leaderboard_file)
return False
except Exception as e:
print(f"Error downloading leaderboard data: {e}")
# Ensure we have at least an empty leaderboard file
leaderboard_file = get_versioned_leaderboard_file(version)
if not os.path.exists(leaderboard_file):
empty_data = {
"entries": [],
"last_updated": pd.Timestamp.now().isoformat(),
"version": version
}
save_leaderboard_data(empty_data, leaderboard_file)
return False
def get_leaderboard_df(version="v0") -> pd.DataFrame:
"""
Get the leaderboard data as a DataFrame.
Args:
version: The dataset version to retrieve
"""
# Try to download the latest data
download_leaderboard_data(version=version)
# Load from local file
leaderboard_file = get_versioned_leaderboard_file(version)
leaderboard_data = load_leaderboard_data(leaderboard_file)
# Convert to DataFrame
df = leaderboard_to_dataframe(leaderboard_data)
return df
def get_category_leaderboard_df(category: str, version="v0") -> pd.DataFrame:
"""
Get the leaderboard data filtered by a specific category.
Args:
category: The category to filter by (e.g., "Criminal, Violent, and Terrorist Activity")
version: The dataset version to retrieve
Returns:
DataFrame with metrics for the specified category
"""
# Load the leaderboard data
leaderboard_file = get_versioned_leaderboard_file(version)
leaderboard_data = load_leaderboard_data(leaderboard_file)
# Filter entries to only include those with data for the specified category
filtered_entries = []
for entry in leaderboard_data.get("entries", []):
# Check if the entry has data for this category
if "per_category_metrics" in entry and category in entry["per_category_metrics"]:
# Create a new entry with just the overall info and this category's metrics
filtered_entry = {
"model_name": entry.get("model_name", "Unknown Model"),
"model_type": entry.get("model_type", "Unknown"),
"submission_date": entry.get("submission_date", ""),
"version": entry.get("version", version),
}
# Extract metrics for this category
category_metrics = entry["per_category_metrics"][category]
# Add metrics for each test type
for test_type in category_metrics:
if test_type and isinstance(category_metrics[test_type], dict):
for metric, value in category_metrics[test_type].items():
col_name = f"{test_type}_{metric}"
filtered_entry[col_name] = value
# Calculate average F1 for this category
f1_values = []
for test_type in category_metrics:
if test_type and isinstance(category_metrics[test_type], dict) and "f1_binary" in category_metrics[test_type]:
f1_values.append(category_metrics[test_type]["f1_binary"])
if f1_values:
filtered_entry["average_f1"] = sum(f1_values) / len(f1_values)
# Add specific test type F1 scores for display
for test_type in ["default_prompts", "jailbreaked_prompts", "default_answers", "jailbreaked_answers"]:
if test_type in category_metrics and "f1_binary" in category_metrics[test_type]:
filtered_entry[f"{test_type}_f1"] = category_metrics[test_type]["f1_binary"]
filtered_entries.append(filtered_entry)
# Create a new leaderboard data structure with the filtered entries
filtered_leaderboard = {
"entries": filtered_entries,
"last_updated": leaderboard_data.get("last_updated", pd.Timestamp.now().isoformat()),
"version": version
}
# Convert to DataFrame
df = leaderboard_to_dataframe(filtered_leaderboard)
return df
def get_detailed_model_data(model_name: str, version="v0") -> Dict:
"""
Get detailed data for a specific model.
Args:
model_name: The name of the model to get data for
version: The dataset version to retrieve
"""
leaderboard_file = get_versioned_leaderboard_file(version)
leaderboard_data = load_leaderboard_data(leaderboard_file)
for entry in leaderboard_data.get("entries", []):
# Check both the model name and version
entry_version = entry.get("version", "v0")
if entry.get("model_name") == model_name and (entry_version == version or entry_version is None):
return entry
return {}
|