Spaces:
Sleeping
Sleeping
Update mmlu_eval_original.py
Browse files- mmlu_eval_original.py +197 -66
mmlu_eval_original.py
CHANGED
@@ -2,10 +2,10 @@ import torch
|
|
2 |
import evaluate
|
3 |
from datasets import load_dataset
|
4 |
from transformers import AutoTokenizer, AutoModelForCausalLM
|
5 |
-
import spaces
|
6 |
import logging
|
7 |
import numpy as np
|
8 |
import pandas as pd
|
|
|
9 |
|
10 |
# Set up logging
|
11 |
logging.basicConfig(level=logging.INFO)
|
@@ -13,7 +13,7 @@ logger = logging.getLogger(__name__)
|
|
13 |
|
14 |
accuracy_metric = evaluate.load("accuracy")
|
15 |
option_letters = ["A", "B", "C", "D"]
|
16 |
-
MAX_CONTEXT_WINDOW = 4096
|
17 |
|
18 |
def load_dataset_from_hf(verbose=False):
|
19 |
mmlu_dataset = load_dataset("cais/mmlu", "all")
|
@@ -93,86 +93,193 @@ def gen_prompt(df, subject, k=-1):
|
|
93 |
|
94 |
|
95 |
@torch.no_grad()
|
96 |
-
def
|
|
|
|
|
|
|
97 |
assert all(dev_df['subject'] == subject), f"Not all items in dev_df match subject {subject}"
|
98 |
assert all(test_df['subject'] == subject), f"Not all items in test_df match subject {subject}"
|
99 |
|
100 |
-
logger.info(f"Subject: {subject}")
|
101 |
|
102 |
cors = []
|
103 |
all_probs = []
|
104 |
|
105 |
if (train_shots < 0):
|
106 |
-
train_shots = 0
|
107 |
-
|
108 |
-
for i in range(test_df.shape[0]):
|
109 |
-
prompt_end = format_example(test_df, i, include_answer=False)
|
110 |
-
train_prompt = gen_prompt(dev_df, subject, train_shots)
|
111 |
-
prompt = train_prompt + prompt_end
|
112 |
|
113 |
-
|
114 |
-
|
115 |
-
|
116 |
-
|
117 |
-
|
118 |
-
|
119 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
120 |
prompt = train_prompt + prompt_end
|
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 |
acc = np.mean(cors)
|
165 |
cors = np.array(cors)
|
166 |
-
|
167 |
all_probs = np.array(all_probs)
|
|
|
168 |
print("Average accuracy {:.3f} - {}".format(acc, subject))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
169 |
|
170 |
-
return cors, acc, all_probs
|
171 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
172 |
|
173 |
-
|
|
|
174 |
"""
|
175 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
176 |
|
177 |
Args:
|
178 |
model: The model to evaluate
|
@@ -180,7 +287,30 @@ def evaluate_mmlu(model, tokenizer, num_subjects=-1, num_questions=5, num_shots=
|
|
180 |
num_subjects (int): Number of subjects to evaluate. If -1, evaluates all subjects
|
181 |
num_questions (int): Number of questions per subject
|
182 |
num_shots (int): Number of few-shot examples to use
|
|
|
|
|
183 |
"""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
184 |
model.eval() # Ensure Dropout and BatchNorm behave appropriately for inference
|
185 |
|
186 |
dataset = load_dataset_from_hf(verbose=True)
|
@@ -207,21 +337,22 @@ def evaluate_mmlu(model, tokenizer, num_subjects=-1, num_questions=5, num_shots=
|
|
207 |
all_cors = []
|
208 |
results_table = []
|
209 |
|
210 |
-
for subject in subjects:
|
211 |
test_samples = test_df[test_df['subject'] == subject].head(num_questions)
|
212 |
dev_samples = dev_df[dev_df['subject'] == subject].head(num_shots)
|
213 |
|
214 |
# Log subject and sample counts
|
215 |
logger.info(f"Subject: {subject}, Test Samples: {len(test_samples)}, Dev Samples: {len(dev_samples)}")
|
216 |
|
217 |
-
cors, acc, probs =
|
218 |
subject,
|
219 |
model,
|
220 |
tokenizer,
|
221 |
dev_samples,
|
222 |
test_samples,
|
223 |
num_questions_per_subject=num_questions,
|
224 |
-
train_shots=num_shots
|
|
|
225 |
)
|
226 |
|
227 |
results[subject] = acc
|
|
|
2 |
import evaluate
|
3 |
from datasets import load_dataset
|
4 |
from transformers import AutoTokenizer, AutoModelForCausalLM
|
|
|
5 |
import logging
|
6 |
import numpy as np
|
7 |
import pandas as pd
|
8 |
+
from tqdm import tqdm
|
9 |
|
10 |
# Set up logging
|
11 |
logging.basicConfig(level=logging.INFO)
|
|
|
13 |
|
14 |
accuracy_metric = evaluate.load("accuracy")
|
15 |
option_letters = ["A", "B", "C", "D"]
|
16 |
+
MAX_CONTEXT_WINDOW = 4096
|
17 |
|
18 |
def load_dataset_from_hf(verbose=False):
|
19 |
mmlu_dataset = load_dataset("cais/mmlu", "all")
|
|
|
93 |
|
94 |
|
95 |
@torch.no_grad()
|
96 |
+
def eval_batched(subject, model, tokenizer, dev_df, test_df, num_questions_per_subject=5, train_shots=5, batch_size=4):
|
97 |
+
"""
|
98 |
+
Improved eval function that uses batched processing on GPU
|
99 |
+
"""
|
100 |
assert all(dev_df['subject'] == subject), f"Not all items in dev_df match subject {subject}"
|
101 |
assert all(test_df['subject'] == subject), f"Not all items in test_df match subject {subject}"
|
102 |
|
103 |
+
logger.info(f"Subject: {subject}, processing with batch_size={batch_size}")
|
104 |
|
105 |
cors = []
|
106 |
all_probs = []
|
107 |
|
108 |
if (train_shots < 0):
|
109 |
+
train_shots = 0 # Make positive.
|
|
|
|
|
|
|
|
|
|
|
110 |
|
111 |
+
# Generate the few-shot examples for this subject once
|
112 |
+
train_prompt = gen_prompt(dev_df, subject, train_shots)
|
113 |
+
|
114 |
+
# Process test examples in batches
|
115 |
+
for batch_start in range(0, test_df.shape[0], batch_size):
|
116 |
+
batch_end = min(batch_start + batch_size, test_df.shape[0])
|
117 |
+
batch_size_actual = batch_end - batch_start
|
118 |
+
|
119 |
+
# Prepare batch prompts
|
120 |
+
batch_prompts = []
|
121 |
+
batch_labels = []
|
122 |
+
|
123 |
+
for i in range(batch_start, batch_end):
|
124 |
+
prompt_end = format_example(test_df, i, include_answer=False)
|
125 |
prompt = train_prompt + prompt_end
|
126 |
+
batch_prompts.append(prompt)
|
127 |
+
|
128 |
+
label = test_df.iloc[i, 3]
|
129 |
+
label_letter = {0: "A", 1: "B", 2: "C", 3: "D"}[label]
|
130 |
+
batch_labels.append(label_letter)
|
131 |
|
132 |
+
# Tokenize all prompts in batch
|
133 |
+
tokenized_inputs = tokenizer(batch_prompts, padding=True, return_tensors="pt")
|
134 |
+
input_ids = tokenized_inputs.input_ids.to(model.device)
|
135 |
+
attention_mask = tokenized_inputs.attention_mask.to(model.device)
|
136 |
+
|
137 |
+
# Check if any example exceeds context window and adjust if needed
|
138 |
+
if input_ids.shape[1] > MAX_CONTEXT_WINDOW:
|
139 |
+
logger.warning(f"Some examples exceed max context window ({input_ids.shape[1]} > {MAX_CONTEXT_WINDOW})")
|
140 |
+
logger.warning(f"Reducing train_shots from {train_shots}")
|
141 |
+
|
142 |
+
# Find the lowest train_shots that fits
|
143 |
+
while train_shots > 0:
|
144 |
+
train_shots -= 1
|
145 |
+
train_prompt = gen_prompt(dev_df, subject, train_shots)
|
146 |
+
|
147 |
+
# Recalculate prompts with fewer shots
|
148 |
+
temp_prompt = train_prompt + format_example(test_df, batch_start, include_answer=False)
|
149 |
+
temp_tokens = tokenizer(temp_prompt, return_tensors="pt").input_ids
|
150 |
+
|
151 |
+
if temp_tokens.shape[1] <= MAX_CONTEXT_WINDOW:
|
152 |
+
logger.info(f"Reduced to train_shots={train_shots}")
|
153 |
+
|
154 |
+
# Regenerate all prompts in the batch with fewer shots
|
155 |
+
batch_prompts = []
|
156 |
+
for i in range(batch_start, batch_end):
|
157 |
+
prompt_end = format_example(test_df, i, include_answer=False)
|
158 |
+
prompt = train_prompt + prompt_end
|
159 |
+
batch_prompts.append(prompt)
|
160 |
+
|
161 |
+
# Retokenize with reduced shots
|
162 |
+
tokenized_inputs = tokenizer(batch_prompts, padding=True, return_tensors="pt")
|
163 |
+
input_ids = tokenized_inputs.input_ids.to(model.device)
|
164 |
+
attention_mask = tokenized_inputs.attention_mask.to(model.device)
|
165 |
+
break
|
166 |
+
|
167 |
+
# If we still can't fit even with 0 shots, we have to skip
|
168 |
+
if input_ids.shape[1] > MAX_CONTEXT_WINDOW:
|
169 |
+
logger.error(f"Even with 0 shots, context is too long ({input_ids.shape[1]} > {MAX_CONTEXT_WINDOW})")
|
170 |
+
# Process individually as fallback
|
171 |
+
for i in range(batch_start, batch_end):
|
172 |
+
single_prompt = format_example(test_df, i, include_answer=False)
|
173 |
+
single_tokens = tokenizer(single_prompt, return_tensors="pt").input_ids.to(model.device)
|
174 |
+
if single_tokens.shape[1] <= MAX_CONTEXT_WINDOW:
|
175 |
+
single_output = model(input_ids=single_tokens)
|
176 |
+
single_logits = single_output.logits[0, -1]
|
177 |
+
single_probs = get_option_probs(tokenizer, single_logits)
|
178 |
+
pred = {0: "A", 1: "B", 2: "C", 3: "D"}[np.argmax(single_probs)]
|
179 |
+
cors.append(pred == batch_labels[i-batch_start])
|
180 |
+
all_probs.append(single_probs)
|
181 |
+
else:
|
182 |
+
logger.error(f"Example {i} is too long even by itself, skipping")
|
183 |
+
continue
|
184 |
+
|
185 |
+
# Run model on batch
|
186 |
+
outputs = model(input_ids=input_ids, attention_mask=attention_mask)
|
187 |
+
|
188 |
+
# Extract predictions for each example in batch
|
189 |
+
for j in range(batch_size_actual):
|
190 |
+
# Get logits for the last token in each sequence
|
191 |
+
sequence_len = attention_mask[j].sum()
|
192 |
+
logits = outputs.logits[j, sequence_len-1]
|
193 |
+
|
194 |
+
# Calculate probabilities for A, B, C, D
|
195 |
+
probs = get_option_probs(tokenizer, logits)
|
196 |
+
pred = {0: "A", 1: "B", 2: "C", 3: "D"}[np.argmax(probs)]
|
197 |
+
|
198 |
+
cor = pred == batch_labels[j]
|
199 |
+
|
200 |
+
# Log first example for debugging
|
201 |
+
if batch_start == 0 and j == 0:
|
202 |
+
logger.info(f"Prompt (truncated): {batch_prompts[j][:200]}...")
|
203 |
+
logger.info(f"Label_Letter: {batch_labels[j]}")
|
204 |
+
logger.info(f"Probabilities: {probs}")
|
205 |
+
logger.info(f"Prediction: {pred}")
|
206 |
+
logger.info(f"Correct: {cor}")
|
207 |
+
|
208 |
+
cors.append(cor)
|
209 |
+
all_probs.append(probs)
|
210 |
+
|
211 |
acc = np.mean(cors)
|
212 |
cors = np.array(cors)
|
|
|
213 |
all_probs = np.array(all_probs)
|
214 |
+
|
215 |
print("Average accuracy {:.3f} - {}".format(acc, subject))
|
216 |
+
|
217 |
+
return subject, cors, acc, all_probs
|
218 |
+
|
219 |
+
|
220 |
+
def get_option_probs(tokenizer, logits):
|
221 |
+
"""Helper function to extract option probabilities from logits"""
|
222 |
+
option_probs = torch.nn.functional.softmax(
|
223 |
+
torch.tensor(
|
224 |
+
[
|
225 |
+
logits[tokenizer("A").input_ids[-1]],
|
226 |
+
logits[tokenizer("B").input_ids[-1]],
|
227 |
+
logits[tokenizer("C").input_ids[-1]],
|
228 |
+
logits[tokenizer("D").input_ids[-1]],
|
229 |
+
]
|
230 |
+
).float(),
|
231 |
+
dim=0,
|
232 |
+
).detach().cpu().numpy()
|
233 |
+
|
234 |
+
return option_probs
|
235 |
|
|
|
236 |
|
237 |
+
def get_max_batch_size(model, tokenizer, example_text, max_memory_fraction=0.8):
|
238 |
+
"""
|
239 |
+
Estimate the maximum possible batch size based on available GPU memory
|
240 |
+
|
241 |
+
Args:
|
242 |
+
model: The model to evaluate
|
243 |
+
tokenizer: The tokenizer to use
|
244 |
+
example_text: A sample text input
|
245 |
+
max_memory_fraction: Maximum fraction of GPU memory to use (0.8 = 80%)
|
246 |
|
247 |
+
Returns:
|
248 |
+
Estimated maximum batch size
|
249 |
"""
|
250 |
+
import torch
|
251 |
+
|
252 |
+
# Get total GPU memory and currently allocated memory
|
253 |
+
total_memory = torch.cuda.get_device_properties(0).total_memory
|
254 |
+
|
255 |
+
# Keep a safe buffer to avoid OOM
|
256 |
+
safe_memory = int(total_memory * max_memory_fraction)
|
257 |
+
|
258 |
+
# Tokenize example to get size
|
259 |
+
example_tokens = tokenizer(example_text, return_tensors="pt").to(model.device)
|
260 |
+
example_len = example_tokens.input_ids.shape[1]
|
261 |
+
|
262 |
+
# Run a single forward pass to measure memory usage
|
263 |
+
torch.cuda.empty_cache()
|
264 |
+
torch.cuda.reset_peak_memory_stats()
|
265 |
+
_ = model(**example_tokens)
|
266 |
+
single_forward_memory = torch.cuda.max_memory_allocated()
|
267 |
+
|
268 |
+
# Calculate memory per example and estimate max batch size
|
269 |
+
estimated_max_batch = safe_memory // single_forward_memory
|
270 |
+
|
271 |
+
# Reduce by a factor for safety (activations, gradients, etc.)
|
272 |
+
safe_batch_size = max(1, int(estimated_max_batch * 0.8))
|
273 |
+
|
274 |
+
logger.info(f"Estimated max batch size: {safe_batch_size} for sequence length {example_len}")
|
275 |
+
logger.info(f"Memory usage: {single_forward_memory / 1e9:.2f} GB per example")
|
276 |
+
logger.info(f"Total memory: {total_memory / 1e9:.2f} GB, Safe memory: {safe_memory / 1e9:.2f} GB")
|
277 |
+
|
278 |
+
return safe_batch_size
|
279 |
+
|
280 |
+
def evaluate_mmlu_batched(model, tokenizer, num_subjects=10, num_questions=10, num_shots=5, batch_size=8, auto_batch_size=False):
|
281 |
+
"""
|
282 |
+
Evaluates the model on MMLU using batched GPU processing for faster inference.
|
283 |
|
284 |
Args:
|
285 |
model: The model to evaluate
|
|
|
287 |
num_subjects (int): Number of subjects to evaluate. If -1, evaluates all subjects
|
288 |
num_questions (int): Number of questions per subject
|
289 |
num_shots (int): Number of few-shot examples to use
|
290 |
+
batch_size (int): Batch size for processing multiple examples at once
|
291 |
+
auto_batch_size (bool): If True, automatically determine the optimal batch size
|
292 |
"""
|
293 |
+
|
294 |
+
# If auto_batch_size is enabled, estimate the optimal batch size
|
295 |
+
if auto_batch_size:
|
296 |
+
# Get a sample prompt
|
297 |
+
dataset = load_dataset_from_hf(verbose=False)
|
298 |
+
test_df = pd.DataFrame(dataset['test'])
|
299 |
+
dev_df = pd.DataFrame(dataset['dev'])
|
300 |
+
test_df = test_df.sort_values(['subject', 'question'])
|
301 |
+
dev_df = dev_df.sort_values(['subject', 'question'])
|
302 |
+
subject = test_df['subject'].iloc[0]
|
303 |
+
test_sample = test_df[test_df['subject'] == subject].head(1)
|
304 |
+
dev_sample = dev_df[dev_df['subject'] == subject].head(num_shots)
|
305 |
+
|
306 |
+
# Generate a sample prompt
|
307 |
+
train_prompt = gen_prompt(dev_sample, subject, num_shots)
|
308 |
+
sample_prompt = train_prompt + format_example(test_sample, 0, include_answer=False)
|
309 |
+
|
310 |
+
# Estimate the max batch size
|
311 |
+
batch_size = get_max_batch_size(model, tokenizer, sample_prompt)
|
312 |
+
logger.info(f"Auto-adjusted batch size: {batch_size}")
|
313 |
+
|
314 |
model.eval() # Ensure Dropout and BatchNorm behave appropriately for inference
|
315 |
|
316 |
dataset = load_dataset_from_hf(verbose=True)
|
|
|
337 |
all_cors = []
|
338 |
results_table = []
|
339 |
|
340 |
+
for subject in tqdm(subjects, desc="Processing subjects"):
|
341 |
test_samples = test_df[test_df['subject'] == subject].head(num_questions)
|
342 |
dev_samples = dev_df[dev_df['subject'] == subject].head(num_shots)
|
343 |
|
344 |
# Log subject and sample counts
|
345 |
logger.info(f"Subject: {subject}, Test Samples: {len(test_samples)}, Dev Samples: {len(dev_samples)}")
|
346 |
|
347 |
+
subject, cors, acc, probs = eval_batched(
|
348 |
subject,
|
349 |
model,
|
350 |
tokenizer,
|
351 |
dev_samples,
|
352 |
test_samples,
|
353 |
num_questions_per_subject=num_questions,
|
354 |
+
train_shots=num_shots,
|
355 |
+
batch_size=batch_size
|
356 |
)
|
357 |
|
358 |
results[subject] = acc
|