update parquet tables and fix typo
Browse files- .gitignore +1 -0
- README.md +2 -2
- boostrap_confidence_intervals-00000-of-00001.parquet → bootstrap_confidence_intervals-00000-of-00001.parquet +2 -2
- create_parquet_files.py +360 -9
- evaluation_results-00000-of-00001.parquet +2 -2
- runs/api_models/compute_bootstrap_ci.py +6 -3
- runs/api_models/metrics.py +6 -3
.gitignore
CHANGED
@@ -1 +1,2 @@
|
|
1 |
runs/api_models/__pycache__/*.pyc
|
|
|
|
1 |
runs/api_models/__pycache__/*.pyc
|
2 |
+
logs/**/*
|
README.md
CHANGED
@@ -8,8 +8,8 @@ configs:
|
|
8 |
path: evaluation_results-*.parquet
|
9 |
- config_name: bootstrap_confidence_intervals
|
10 |
data_files:
|
11 |
-
- split:
|
12 |
-
path:
|
13 |
tags:
|
14 |
- automatic-essay-scoring
|
15 |
- portuguese
|
|
|
8 |
path: evaluation_results-*.parquet
|
9 |
- config_name: bootstrap_confidence_intervals
|
10 |
data_files:
|
11 |
+
- split: bootstrap_confidence_intervals
|
12 |
+
path: bootstrap_confidence_intervals-*.parquet
|
13 |
tags:
|
14 |
- automatic-essay-scoring
|
15 |
- portuguese
|
boostrap_confidence_intervals-00000-of-00001.parquet → bootstrap_confidence_intervals-00000-of-00001.parquet
RENAMED
@@ -1,3 +1,3 @@
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:
|
3 |
-
size
|
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:90fa615f5ff10a5ff533b5c8e65df895ba313844d9e5d708573f7d3e81787fc4
|
3 |
+
size 28166
|
create_parquet_files.py
CHANGED
@@ -1,40 +1,391 @@
|
|
1 |
import pandas as pd
|
2 |
from pathlib import Path
|
3 |
import pyarrow # ensures pyarrow is installed for Parquet support
|
|
|
|
|
|
|
|
|
|
|
4 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
5 |
|
6 |
def find_and_group_csvs():
|
7 |
base = Path(".")
|
8 |
groups = {
|
9 |
"evaluation_results": sorted(base.rglob("evaluation_results.csv")),
|
10 |
-
"bootstrap_confidence_intervals": sorted(
|
|
|
|
|
11 |
}
|
12 |
for name, paths in groups.items():
|
13 |
-
|
14 |
if not paths:
|
15 |
-
|
16 |
return groups
|
17 |
|
18 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
19 |
def combine(paths, out_path):
|
20 |
if not paths:
|
21 |
-
|
22 |
return
|
23 |
|
24 |
-
|
25 |
-
dfs = [
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
26 |
|
27 |
# Basic schema validation
|
28 |
cols = {tuple(df.columns) for df in dfs}
|
29 |
if len(cols) > 1:
|
30 |
-
|
|
|
|
|
31 |
|
32 |
combined = pd.concat(dfs, ignore_index=True)
|
|
|
|
|
|
|
|
|
|
|
|
|
33 |
combined.to_parquet(out_path, engine="pyarrow", index=False)
|
34 |
-
|
35 |
|
36 |
|
37 |
if __name__ == "__main__":
|
|
|
|
|
|
|
38 |
groups = find_and_group_csvs()
|
39 |
combine(groups["evaluation_results"], "evaluation_results-00000-of-00001.parquet")
|
40 |
-
combine(
|
|
|
|
|
|
|
|
|
|
|
|
1 |
import pandas as pd
|
2 |
from pathlib import Path
|
3 |
import pyarrow # ensures pyarrow is installed for Parquet support
|
4 |
+
import numpy as np
|
5 |
+
import sys
|
6 |
+
from tqdm.auto import tqdm
|
7 |
+
import logging
|
8 |
+
from datetime import datetime
|
9 |
|
10 |
+
# Add the api_models directory to the Python path to import existing modules
|
11 |
+
sys.path.append(str(Path(__file__).parent / "runs" / "api_models"))
|
12 |
+
|
13 |
+
from compute_bootstrap_ci import (
|
14 |
+
load_inference_results_by_grader,
|
15 |
+
extract_config_from_log,
|
16 |
+
)
|
17 |
+
from metrics import compute_metrics
|
18 |
+
from omegaconf import OmegaConf
|
19 |
+
|
20 |
+
# Set up logging
|
21 |
+
log_dir = Path("logs")
|
22 |
+
log_dir.mkdir(exist_ok=True)
|
23 |
+
log_file = log_dir / f"create_parquet_files_{datetime.now().strftime('%Y%m%d_%H%M%S')}.log"
|
24 |
+
|
25 |
+
# Configure logging to write to both file and console
|
26 |
+
logging.basicConfig(
|
27 |
+
level=logging.INFO,
|
28 |
+
format='%(asctime)s - %(levelname)s - %(message)s',
|
29 |
+
handlers=[
|
30 |
+
logging.FileHandler(log_file),
|
31 |
+
logging.StreamHandler(sys.stdout)
|
32 |
+
]
|
33 |
+
)
|
34 |
+
logger = logging.getLogger(__name__)
|
35 |
+
|
36 |
+
# Also create a separate error-only log file
|
37 |
+
error_log_file = log_dir / f"create_parquet_files_errors_{datetime.now().strftime('%Y%m%d_%H%M%S')}.log"
|
38 |
+
error_handler = logging.FileHandler(error_log_file)
|
39 |
+
error_handler.setLevel(logging.ERROR)
|
40 |
+
error_handler.setFormatter(logging.Formatter('%(asctime)s - %(levelname)s - %(message)s'))
|
41 |
+
logger.addHandler(error_handler)
|
42 |
+
|
43 |
+
def simplify_experiment_name(name):
|
44 |
+
"""Simplify experiment names according to the mapping rules."""
|
45 |
+
if pd.isna(name):
|
46 |
+
return name
|
47 |
+
|
48 |
+
# Convert to string to handle any non-string inputs
|
49 |
+
name = str(name)
|
50 |
+
|
51 |
+
# Define the mapping rules
|
52 |
+
mappings = {
|
53 |
+
# Sabia-3 mappings
|
54 |
+
'sabia-3-zero-shot': 'sabia3-studentPrompt',
|
55 |
+
'sabia-3-extractor-zero-shot': 'sabia3-extractor',
|
56 |
+
'sabia-3-grader-zero-shot': 'sabia3-graderPrompt',
|
57 |
+
|
58 |
+
# Deepseek mappings
|
59 |
+
'deepseek-reasoner-zero-shot': 'deepseekR1-studentPrompt',
|
60 |
+
'deepseek-reasoner-extractor-zero-shot': 'deepseekR1-extractor',
|
61 |
+
'deepseek-reasoner-grader-zero-shot': 'deepseekR1-graderPrompt',
|
62 |
+
|
63 |
+
# GPT-4o mappings
|
64 |
+
'gpt-4o-2024-11-20-zero-shot': 'gpt4o-studentPrompt',
|
65 |
+
'gpt-4o-2024-11-20-extractor-zero-shot': 'gpt4o-extractor',
|
66 |
+
'gpt-4o-2024-11-20-grader-zero-shot': 'gpt4o-graderPrompt',
|
67 |
+
}
|
68 |
+
|
69 |
+
# Apply direct mappings first
|
70 |
+
for pattern, replacement in mappings.items():
|
71 |
+
if pattern in name:
|
72 |
+
name = name.replace(pattern, replacement)
|
73 |
+
|
74 |
+
# Handle jbcs2025 prefixed names
|
75 |
+
if name.startswith('jbcs2025_'):
|
76 |
+
# Remove the prefix
|
77 |
+
name = name[9:]
|
78 |
+
|
79 |
+
# First, remove any duplicated model-specific patterns that appear multiple times
|
80 |
+
# These patterns indicate the experiment setup was duplicated in the name
|
81 |
+
duplication_patterns = [
|
82 |
+
'llama31_classification_lora',
|
83 |
+
'phi35_classification_lora',
|
84 |
+
'phi4_classification_lora',
|
85 |
+
'encoder_classification'
|
86 |
+
]
|
87 |
+
|
88 |
+
for pattern in duplication_patterns:
|
89 |
+
# Count occurrences
|
90 |
+
count = name.count(f'-{pattern}-')
|
91 |
+
if count > 1:
|
92 |
+
# Replace all but keep track of components
|
93 |
+
parts = name.split(f'-{pattern}-')
|
94 |
+
# Keep the first part and the last part (which has the config)
|
95 |
+
if len(parts) > 2:
|
96 |
+
name = parts[0] + '-' + parts[-1]
|
97 |
+
|
98 |
+
# Handle BERT variants
|
99 |
+
if 'bert-base-portuguese-cased-encoder_classification' in name:
|
100 |
+
name = name.replace('bert-base-portuguese-cased-encoder_classification', 'bertimbau-base')
|
101 |
+
elif 'BERTugues-base-portuguese-cased-encoder_classification' in name:
|
102 |
+
name = name.replace('BERTugues-base-portuguese-cased-encoder_classification', 'bertugues-base')
|
103 |
+
elif 'bert-base-multilingual-cased-encoder_classification' in name:
|
104 |
+
name = name.replace('bert-base-multilingual-cased-encoder_classification', 'mbert-base')
|
105 |
+
elif 'bert-large-portuguese-cased-encoder_classification' in name:
|
106 |
+
name = name.replace('bert-large-portuguese-cased-encoder_classification', 'bertimbau-large')
|
107 |
+
|
108 |
+
# Handle Llama variants
|
109 |
+
elif 'Llama-3.1-8B-llama31_classification_lora' in name:
|
110 |
+
name = name.replace('Llama-3.1-8B-llama31_classification_lora', 'llama3.1-8b-lora')
|
111 |
+
elif 'Llama-3.1-8B' in name:
|
112 |
+
name = name.replace('Llama-3.1-8B', 'llama3.1-8b-lora')
|
113 |
+
|
114 |
+
# Handle Phi variants
|
115 |
+
elif 'Phi-3.5-mini-instruct-phi35_classification_lora' in name:
|
116 |
+
name = name.replace('Phi-3.5-mini-instruct-phi35_classification_lora', 'phi3.5-mini-lora')
|
117 |
+
elif 'Phi-3.5-mini-instruct' in name:
|
118 |
+
name = name.replace('Phi-3.5-mini-instruct', 'phi3.5-mini-lora')
|
119 |
+
elif 'phi-4-phi4_classification_lora' in name:
|
120 |
+
name = name.replace('phi-4-phi4_classification_lora', 'phi4-lora')
|
121 |
+
elif 'phi-4' in name:
|
122 |
+
name = name.replace('phi-4', 'phi4-lora')
|
123 |
+
|
124 |
+
# Clean up any remaining classification patterns
|
125 |
+
name = name.replace('-encoder_classification', '')
|
126 |
+
name = name.replace('_classification_lora', '')
|
127 |
+
name = name.replace('-llama31', '')
|
128 |
+
name = name.replace('-phi35', '')
|
129 |
+
name = name.replace('-phi4', '')
|
130 |
+
|
131 |
+
# Extract components and reorder
|
132 |
+
parts = name.split('-')
|
133 |
+
|
134 |
+
# Look for competency (C1-C5), context type, and LoRA rank
|
135 |
+
competency = None
|
136 |
+
context = None
|
137 |
+
lora_rank = None
|
138 |
+
model_parts = []
|
139 |
+
|
140 |
+
i = 0
|
141 |
+
while i < len(parts):
|
142 |
+
part = parts[i]
|
143 |
+
if part in ['C1', 'C2', 'C3', 'C4', 'C5']:
|
144 |
+
competency = part
|
145 |
+
elif part == 'essay_only':
|
146 |
+
context = 'essay-only'
|
147 |
+
elif part == 'full_context':
|
148 |
+
context = 'full-context'
|
149 |
+
elif part in ['essay', 'full'] and i + 1 < len(parts):
|
150 |
+
# Handle split context names
|
151 |
+
if parts[i+1] == 'only':
|
152 |
+
context = 'essay-only'
|
153 |
+
i += 1 # Skip next part
|
154 |
+
elif parts[i+1] == 'context':
|
155 |
+
context = 'full-context'
|
156 |
+
i += 1 # Skip next part
|
157 |
+
elif part in ['r8', 'r16']:
|
158 |
+
lora_rank = part
|
159 |
+
elif part and part not in ['only', 'context']: # Skip empty parts and orphaned context words
|
160 |
+
model_parts.append(part)
|
161 |
+
i += 1
|
162 |
+
|
163 |
+
# Reconstruct the name in the desired order: model-competency-context-rank
|
164 |
+
new_parts = model_parts
|
165 |
+
if competency:
|
166 |
+
new_parts.append(competency)
|
167 |
+
if context:
|
168 |
+
new_parts.append(context)
|
169 |
+
if lora_rank:
|
170 |
+
new_parts.append(lora_rank)
|
171 |
+
|
172 |
+
name = '-'.join(new_parts)
|
173 |
+
|
174 |
+
# Final cleanup: remove any double dashes
|
175 |
+
while '--' in name:
|
176 |
+
name = name.replace('--', '-')
|
177 |
+
|
178 |
+
return name
|
179 |
|
180 |
def find_and_group_csvs():
|
181 |
base = Path(".")
|
182 |
groups = {
|
183 |
"evaluation_results": sorted(base.rglob("evaluation_results.csv")),
|
184 |
+
"bootstrap_confidence_intervals": sorted(
|
185 |
+
base.rglob("bootstrap_confidence_intervals.csv")
|
186 |
+
),
|
187 |
}
|
188 |
for name, paths in groups.items():
|
189 |
+
logger.info(f"Found {len(paths)} files for '{name}'")
|
190 |
if not paths:
|
191 |
+
logger.warning(f"No files found for '{name}'")
|
192 |
return groups
|
193 |
|
194 |
|
195 |
+
def enhance_evaluation_results(eval_df, csv_paths):
|
196 |
+
"""Enhance evaluation results with additional metrics from JSONL files."""
|
197 |
+
enhanced_rows = []
|
198 |
+
failed_count = 0
|
199 |
+
|
200 |
+
# Create a mapping from row index to CSV path
|
201 |
+
# Since we're processing multiple CSVs that get concatenated,
|
202 |
+
# we need to track which rows came from which CSV file
|
203 |
+
row_to_path = {}
|
204 |
+
current_idx = 0
|
205 |
+
|
206 |
+
for path in csv_paths:
|
207 |
+
df = pd.read_csv(path)
|
208 |
+
for i in range(len(df)):
|
209 |
+
row_to_path[current_idx + i] = path
|
210 |
+
current_idx += len(df)
|
211 |
+
|
212 |
+
for idx, row in tqdm(
|
213 |
+
eval_df.iterrows(), desc="Processing evaluation rows", total=len(eval_df)
|
214 |
+
):
|
215 |
+
# Get the CSV path for this row
|
216 |
+
csv_path = row_to_path.get(idx)
|
217 |
+
|
218 |
+
if csv_path is None:
|
219 |
+
error_msg = f"CSV file not found for row {idx}"
|
220 |
+
logger.error(error_msg)
|
221 |
+
failed_count += 1
|
222 |
+
continue
|
223 |
+
|
224 |
+
try:
|
225 |
+
# Extract experiment ID from the path
|
226 |
+
# The experiment ID is typically the parent directory name
|
227 |
+
experiment_id = csv_path.parent.name
|
228 |
+
|
229 |
+
# Simplify the experiment ID
|
230 |
+
experiment_id = simplify_experiment_name(experiment_id)
|
231 |
+
|
232 |
+
# Find corresponding JSONL file in the same directory
|
233 |
+
jsonl_path = csv_path.parent / "inference_results.jsonl"
|
234 |
+
if not jsonl_path.exists():
|
235 |
+
# Try with experiment name prefix
|
236 |
+
jsonl_files = list(csv_path.parent.glob("*_inference_results.jsonl"))
|
237 |
+
if jsonl_files:
|
238 |
+
jsonl_path = jsonl_files[0]
|
239 |
+
else:
|
240 |
+
raise FileNotFoundError(f"JSONL file not found in {csv_path.parent}")
|
241 |
+
|
242 |
+
# Find log file to extract configuration
|
243 |
+
log_files = list(csv_path.parent.glob("*run_inference_experiment.log"))
|
244 |
+
if not log_files:
|
245 |
+
raise FileNotFoundError(f"Log file not found in {csv_path.parent}")
|
246 |
+
|
247 |
+
log_path = log_files[0]
|
248 |
+
|
249 |
+
# Load inference results and compute metrics
|
250 |
+
# Extract configuration from log file
|
251 |
+
config_dict = extract_config_from_log(log_path)
|
252 |
+
# Convert to OmegaConf DictConfig for compatibility with compute_metrics
|
253 |
+
cfg = OmegaConf.create(config_dict)
|
254 |
+
|
255 |
+
# Load data using the existing function
|
256 |
+
grader_a_data, grader_b_data = load_inference_results_by_grader(jsonl_path)
|
257 |
+
|
258 |
+
# Extract predictions and labels for each grader
|
259 |
+
all_predictions_a = np.array(
|
260 |
+
[data["prediction"] for data in grader_a_data.values()]
|
261 |
+
)
|
262 |
+
all_labels_a = np.array([data["label"] for data in grader_a_data.values()])
|
263 |
+
all_predictions_b = np.array(
|
264 |
+
[data["prediction"] for data in grader_b_data.values()]
|
265 |
+
)
|
266 |
+
all_labels_b = np.array([data["label"] for data in grader_b_data.values()])
|
267 |
+
|
268 |
+
# Compute concat(A,B) metrics for verification
|
269 |
+
# Concatenate predictions and labels from both graders
|
270 |
+
concat_predictions = np.concatenate([all_predictions_a, all_predictions_b])
|
271 |
+
concat_labels = np.concatenate([all_labels_a, all_labels_b])
|
272 |
+
metrics_concat = compute_metrics((concat_predictions, concat_labels), cfg)
|
273 |
+
|
274 |
+
# Verify that computed concat metrics match original CSV values
|
275 |
+
# Check a few key metrics with some tolerance for floating point comparison
|
276 |
+
tolerance = 1e-6
|
277 |
+
for metric in ["accuracy", "QWK", "Macro_F1", "Weighted_F1"]:
|
278 |
+
if metric in row and metric in metrics_concat:
|
279 |
+
original_value = row[metric]
|
280 |
+
computed_value = metrics_concat[metric]
|
281 |
+
# You can make this a hard assertion if needed:
|
282 |
+
assert abs(original_value - computed_value) <= tolerance, (
|
283 |
+
f"Metric {metric} mismatch: CSV={original_value}, Computed={computed_value}"
|
284 |
+
)
|
285 |
+
|
286 |
+
# 1. Add original row with concat(A,B) metrics
|
287 |
+
concat_row = row.copy()
|
288 |
+
concat_row["experiment_id"] = experiment_id
|
289 |
+
concat_row["metric_group"] = "concat(A,B)"
|
290 |
+
enhanced_rows.append(concat_row)
|
291 |
+
|
292 |
+
# 2. Compute metrics for A and B separately first
|
293 |
+
metrics_a = compute_metrics((all_predictions_a, all_labels_a), cfg)
|
294 |
+
metrics_b = compute_metrics((all_predictions_b, all_labels_b), cfg)
|
295 |
+
|
296 |
+
# 3. Compute avg(A,B) as the average of metrics, not metrics of averaged predictions
|
297 |
+
avg_row = row.copy()
|
298 |
+
avg_row["experiment_id"] = experiment_id
|
299 |
+
avg_row["metric_group"] = "avg(A,B)"
|
300 |
+
# Average the metrics from A and B
|
301 |
+
for metric in metrics_a:
|
302 |
+
if metric in metrics_b and metric in avg_row:
|
303 |
+
avg_value = (metrics_a[metric] + metrics_b[metric]) / 2
|
304 |
+
avg_row[metric] = avg_value
|
305 |
+
enhanced_rows.append(avg_row)
|
306 |
+
|
307 |
+
# 4. Add onlyA metrics
|
308 |
+
only_a_row = row.copy()
|
309 |
+
only_a_row["experiment_id"] = experiment_id
|
310 |
+
only_a_row["metric_group"] = "onlyA"
|
311 |
+
# Update metric columns with onlyA values
|
312 |
+
for metric, value in metrics_a.items():
|
313 |
+
if metric in only_a_row:
|
314 |
+
only_a_row[metric] = value
|
315 |
+
enhanced_rows.append(only_a_row)
|
316 |
+
|
317 |
+
# 5. Add onlyB metrics
|
318 |
+
only_b_row = row.copy()
|
319 |
+
only_b_row["experiment_id"] = experiment_id
|
320 |
+
only_b_row["metric_group"] = "onlyB"
|
321 |
+
# Update metric columns with onlyB values
|
322 |
+
for metric, value in metrics_b.items():
|
323 |
+
if metric in only_b_row:
|
324 |
+
only_b_row[metric] = value
|
325 |
+
enhanced_rows.append(only_b_row)
|
326 |
+
|
327 |
+
except Exception as e:
|
328 |
+
failed_count += 1
|
329 |
+
error_msg = f"Failed to process {csv_path.parent if csv_path else 'unknown path'}: {str(e)}"
|
330 |
+
logger.error(error_msg)
|
331 |
+
# Log full traceback for debugging
|
332 |
+
import traceback
|
333 |
+
logger.error(f"Traceback:\n{traceback.format_exc()}")
|
334 |
+
# Skip this row and continue with the next one
|
335 |
+
continue
|
336 |
+
|
337 |
+
logger.info(f"Successfully processed {len(enhanced_rows)//4} out of {len(eval_df)} rows")
|
338 |
+
if failed_count > 0:
|
339 |
+
logger.warning(f"Failed to process {failed_count} rows. Check error log: {error_log_file}")
|
340 |
+
|
341 |
+
return pd.DataFrame(enhanced_rows)
|
342 |
+
|
343 |
+
|
344 |
def combine(paths, out_path):
|
345 |
if not paths:
|
346 |
+
logger.info(f"No files to combine for {out_path}")
|
347 |
return
|
348 |
|
349 |
+
logger.info(f"Combining {len(paths)} files into {out_path}")
|
350 |
+
dfs = []
|
351 |
+
|
352 |
+
for p in paths:
|
353 |
+
df = pd.read_csv(p)
|
354 |
+
|
355 |
+
# Add experiment_id column based on the parent directory name
|
356 |
+
experiment_id = p.parent.name
|
357 |
+
experiment_id = simplify_experiment_name(experiment_id)
|
358 |
+
df["experiment_id"] = experiment_id
|
359 |
+
|
360 |
+
dfs.append(df)
|
361 |
|
362 |
# Basic schema validation
|
363 |
cols = {tuple(df.columns) for df in dfs}
|
364 |
if len(cols) > 1:
|
365 |
+
error_msg = f"{out_path}: header mismatch across shards"
|
366 |
+
logger.error(error_msg)
|
367 |
+
raise ValueError(error_msg)
|
368 |
|
369 |
combined = pd.concat(dfs, ignore_index=True)
|
370 |
+
|
371 |
+
# Enhance evaluation results with additional metrics
|
372 |
+
if "evaluation_results" in out_path:
|
373 |
+
logger.info("Enhancing evaluation results with additional metrics...")
|
374 |
+
combined = enhance_evaluation_results(combined, paths)
|
375 |
+
|
376 |
combined.to_parquet(out_path, engine="pyarrow", index=False)
|
377 |
+
logger.info(f"Successfully written {out_path} with {len(combined)} rows")
|
378 |
|
379 |
|
380 |
if __name__ == "__main__":
|
381 |
+
logger.info(f"Starting parquet file creation. Logs will be saved to: {log_file}")
|
382 |
+
logger.info(f"Error-only log will be saved to: {error_log_file}")
|
383 |
+
|
384 |
groups = find_and_group_csvs()
|
385 |
combine(groups["evaluation_results"], "evaluation_results-00000-of-00001.parquet")
|
386 |
+
combine(
|
387 |
+
groups["bootstrap_confidence_intervals"],
|
388 |
+
"bootstrap_confidence_intervals-00000-of-00001.parquet",
|
389 |
+
)
|
390 |
+
|
391 |
+
logger.info("Parquet file creation completed")
|
evaluation_results-00000-of-00001.parquet
CHANGED
@@ -1,3 +1,3 @@
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:
|
3 |
-
size
|
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:31558c69a4e80b9e1dac959968c45fefec8401758dec2bb47732f374908a0960
|
3 |
+
size 65235
|
runs/api_models/compute_bootstrap_ci.py
CHANGED
@@ -16,7 +16,7 @@ def extract_config_from_log(log_path: Path) -> Dict:
|
|
16 |
Parse the experiment configuration from run_experiment.log file.
|
17 |
The config is in YAML format at the beginning of the log file.
|
18 |
"""
|
19 |
-
with open(log_path, 'r') as f:
|
20 |
lines = f.readlines()
|
21 |
|
22 |
# Find the start of the YAML config (after the first log line)
|
@@ -58,9 +58,12 @@ def load_inference_results_by_grader(jsonl_path: Path) -> Tuple[Dict[str, Dict],
|
|
58 |
for line in f:
|
59 |
data = json.loads(line.strip())
|
60 |
essay_id = (data['id'], data['id_prompt'], data['essay_text'])
|
61 |
-
|
|
|
|
|
|
|
62 |
essay_data = {
|
63 |
-
'prediction':
|
64 |
'label': data['label']
|
65 |
}
|
66 |
|
|
|
16 |
Parse the experiment configuration from run_experiment.log file.
|
17 |
The config is in YAML format at the beginning of the log file.
|
18 |
"""
|
19 |
+
with open(log_path, 'r', encoding="latin1") as f:
|
20 |
lines = f.readlines()
|
21 |
|
22 |
# Find the start of the YAML config (after the first log line)
|
|
|
58 |
for line in f:
|
59 |
data = json.loads(line.strip())
|
60 |
essay_id = (data['id'], data['id_prompt'], data['essay_text'])
|
61 |
+
# Determine prediction field based on model type
|
62 |
+
model_types = ["gpt", "sabia", "deepseek"]
|
63 |
+
prediction_field = "pontuacao" if any(model in jsonl_path.name for model in model_types) else "prediction"
|
64 |
+
prediction = data[prediction_field]
|
65 |
essay_data = {
|
66 |
+
'prediction': prediction,
|
67 |
'label': data['label']
|
68 |
}
|
69 |
|
runs/api_models/metrics.py
CHANGED
@@ -57,12 +57,15 @@ def _process_predictions(eval_pred, model_type: str) -> Tuple[List[int], List[in
|
|
57 |
|
58 |
elif _is_classification_model(model_type):
|
59 |
# Classification and ordinal models return logits
|
60 |
-
|
61 |
|
62 |
# Ensure true labels are in the correct format (original scale)
|
63 |
if isinstance(all_true_labels[0], (int, np.integer)) and max(all_true_labels) <= 5:
|
64 |
all_true_labels = all_true_labels * 40
|
65 |
-
|
|
|
|
|
|
|
66 |
return all_predictions.tolist(), all_true_labels.tolist()
|
67 |
|
68 |
else:
|
@@ -73,9 +76,9 @@ def compute_metrics(eval_pred, cfg):
|
|
73 |
"""Compute evaluation metrics for the model."""
|
74 |
transformers_logger = logging.getLogger("transformers")
|
75 |
model_type = cfg.experiments.model.type
|
76 |
-
|
77 |
try:
|
78 |
# Process predictions based on model type
|
|
|
79 |
all_predictions, all_true_labels = _process_predictions(eval_pred, model_type)
|
80 |
|
81 |
# Compute metrics
|
|
|
57 |
|
58 |
elif _is_classification_model(model_type):
|
59 |
# Classification and ordinal models return logits
|
60 |
+
all_predictions, all_true_labels = eval_pred
|
61 |
|
62 |
# Ensure true labels are in the correct format (original scale)
|
63 |
if isinstance(all_true_labels[0], (int, np.integer)) and max(all_true_labels) <= 5:
|
64 |
all_true_labels = all_true_labels * 40
|
65 |
+
# Ensure true labels are in the correct format (original scale)
|
66 |
+
if isinstance(all_predictions[0], (int, np.integer)) and max(all_predictions) <= 5:
|
67 |
+
all_predictions = all_predictions * 40
|
68 |
+
|
69 |
return all_predictions.tolist(), all_true_labels.tolist()
|
70 |
|
71 |
else:
|
|
|
76 |
"""Compute evaluation metrics for the model."""
|
77 |
transformers_logger = logging.getLogger("transformers")
|
78 |
model_type = cfg.experiments.model.type
|
|
|
79 |
try:
|
80 |
# Process predictions based on model type
|
81 |
+
|
82 |
all_predictions, all_true_labels = _process_predictions(eval_pred, model_type)
|
83 |
|
84 |
# Compute metrics
|