Spaces:
Sleeping
Sleeping
Update mmlu_eval_original.py
Browse files- mmlu_eval_original.py +36 -2
mmlu_eval_original.py
CHANGED
@@ -3,9 +3,42 @@ import evaluate
|
|
3 |
from datasets import load_dataset
|
4 |
from transformers import AutoTokenizer, AutoModelForCausalLM
|
5 |
import spaces
|
|
|
|
|
|
|
|
|
|
|
6 |
|
7 |
accuracy_metric = evaluate.load("accuracy")
|
8 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
9 |
|
10 |
def format_mmlu_prompt(question, choices):
|
11 |
"""
|
@@ -46,10 +79,11 @@ def generate_answer(model, tokenizer, question, choices):
|
|
46 |
return response[:1] # Fallback: take first character
|
47 |
|
48 |
@torch.no_grad()
|
49 |
-
def
|
50 |
"""
|
51 |
Evaluates the model on MMLU across all tasks.
|
52 |
"""
|
|
|
53 |
results = {}
|
54 |
correct_examples = []
|
55 |
incorrect_examples = []
|
|
|
3 |
from datasets import load_dataset
|
4 |
from transformers import AutoTokenizer, AutoModelForCausalLM
|
5 |
import spaces
|
6 |
+
import logging
|
7 |
+
|
8 |
+
# Set up logging
|
9 |
+
logging.basicConfig(level=logging.INFO)
|
10 |
+
logger = logging.getLogger(__name__)
|
11 |
|
12 |
accuracy_metric = evaluate.load("accuracy")
|
13 |
+
|
14 |
+
def load_dataset_from_hf(verbose=False):
|
15 |
+
mmlu_dataset = load_dataset("cais/mmlu", "all")
|
16 |
+
if verbose:
|
17 |
+
for split in mmlu_dataset.keys():
|
18 |
+
dataset = mmlu_dataset[split] # Access the dataset split
|
19 |
+
|
20 |
+
# Log number of rows and columns
|
21 |
+
num_rows = len(dataset)
|
22 |
+
num_cols = len(dataset.column_names)
|
23 |
+
|
24 |
+
logger.info(f"Dataset Split: {split}")
|
25 |
+
logger.info(f"Number of Rows: {num_rows}")
|
26 |
+
logger.info(f"Number of Columns: {num_cols}")
|
27 |
+
|
28 |
+
# Log column names and their types
|
29 |
+
column_types = {col: str(dataset.features[col].dtype) for col in dataset.column_names}
|
30 |
+
logger.info(f"Column Names: {dataset.column_names}")
|
31 |
+
logger.info(f"Column Types: {column_types}")
|
32 |
+
|
33 |
+
# Log a sample of 5 rows
|
34 |
+
sample_rows = dataset.select(range(min(5, num_rows))) # Ensure we don't exceed available rows
|
35 |
+
logger.info("Sample Rows:")
|
36 |
+
for row in sample_rows:
|
37 |
+
logger.info(row)
|
38 |
+
|
39 |
+
logger.info("=" * 50) # Separator for readability
|
40 |
+
return mmlu_dataset
|
41 |
+
|
42 |
|
43 |
def format_mmlu_prompt(question, choices):
|
44 |
"""
|
|
|
79 |
return response[:1] # Fallback: take first character
|
80 |
|
81 |
@torch.no_grad()
|
82 |
+
def evaluate_mmlu(model, tokenizer, num_questions=5):
|
83 |
"""
|
84 |
Evaluates the model on MMLU across all tasks.
|
85 |
"""
|
86 |
+
mmlu_dataset = load_dataset_from_hf(verbose=True)
|
87 |
results = {}
|
88 |
correct_examples = []
|
89 |
incorrect_examples = []
|