File size: 4,104 Bytes
3608705 | 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 | import ast
from datasets import load_dataset
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
print("1. Loading LogistikaBench from Hugging Face...")
# Replace with your actual Hugging Face repo path and file name if different
dataset = load_dataset("berdymurad/LogistikaBench", data_files="logistikabench.csv")
test_data = dataset["train"]
print(f"Loaded {len(test_data)} evaluation items successfully!")
print("\n2. Loading test model (Qwen2.5-1.5B-Instruct)...")
model_id = "Qwen/Qwen2-7B-Instruct" # Changed model_id to a more powerful model
tokenizer = AutoTokenizer.from_pretrained(model_id)
try:
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype=torch.float16,
device_map="auto"
)
except Exception as e:
print(f"Error loading model: {e}")
raise # Re-raise the exception after printing for full traceback
correct_predictions = 0
total_evaluated = 0
print("\n3. Starting benchmark loop...")
for i, item in enumerate(test_data):
question = item.get('question', '')
# Safely parse choices and answers from string representation (e.g., "['Choice A', 'Choice B']")
raw_choices = item.get('choices', '[]')
if isinstance(raw_choices, str):
choices_list = ast.literal_eval(raw_choices)
else:
choices_list = raw_choices
raw_answer = item.get('answer', '[]')
if isinstance(raw_answer, str):
true_indices = ast.literal_eval(raw_answer)
else:
true_indices = raw_answer
# Map zero-based indices [0, 1, 2] to letters ['A', 'B', 'C']
letter_mapping = {idx: chr(65 + idx) for idx in range(len(choices_list))}
true_letters = set([letter_mapping[idx] for idx in true_indices if idx in letter_mapping])
# Format choices into a readable multiple-choice block (A. Choice text...)
formatted_choices = "\n".join([f"{letter_mapping[idx]}: {choice}" for idx, choice in enumerate(choices_list)])
# Construct prompt instructing the model to output choice letters
prompt = (
f"Answer the following logistics and supply chain multiple-choice question. "
f"Note that there may be ONE or MORE correct answers. "
f"Provide the letters of all correct options together (e.g., A, or AB, or A, C).\n\n"
f"Question: {question}\n\n"
f"Choices:\n{formatted_choices}\n\n"
f"Answers (give only the letters):"
)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
**inputs,
max_new_tokens=15,
temperature=0.0,
do_sample=False
)
response = tokenizer.decode(outputs[0][inputs['input_ids'].shape[1]:], skip_special_tokens=True).upper().strip()
# Define the set of valid answer letters for this specific question
valid_choice_letters = set(letter_mapping.values())
# Extract predicted uppercase letters from the response, strictly from the beginning
# and only until a non-answer-related character is encountered.
predicted_letters_temp = set()
for char in response:
if char in valid_choice_letters:
predicted_letters_temp.add(char)
elif char.isspace() or char == ',':
# Allow spaces and commas as separators within the answer sequence
continue
else:
# Stop if any other character is encountered, assuming the answer sequence has ended
break
predicted_letters = predicted_letters_temp
# Strict match check (model must pick the exact set of correct letters)
if predicted_letters == true_letters:
correct_predictions += 1
total_evaluated += 1
# Debug print statements removed
if (i + 1) % 50 == 0:
print(f"Processed {i + 1} / {len(test_data)} questions...")
# Final Score Calculation
accuracy = (correct_predictions / total_evaluated) * 100
print("\n================================")
print(f"🏁 BENCHMARK COMPLETE!")
print(f"Total Questions Evaluated: {total_evaluated}")
print(f"Strict Match Accuracy: {accuracy:.2f}%")
print("================================") |