BLEnD / evaluation /multiple_choice_evaluation.py
nayeon212's picture
ADD: Data & Code
605bdce verified
from evaluation_utils import *
from multiple_choice_generation import *
def get_model_mc_response(model_name,model_cache_dir,mc_dir,questions_file,response_file=None,temperature=1,top_p=0,gpt_azure=True):
if response_file == None:
response_file = f"{model_name}-mc_res.csv"
questions_df = pd.read_csv(os.path.join(mc_dir,questions_file),encoding='utf-8')
already = None
if not os.path.exists(os.path.join(mc_dir,response_file)):
write_csv_row(list(questions_df.columns)+['full_res','final_ans'],os.path.join(mc_dir,response_file))
else:
already = pd.read_csv(os.path.join(mc_dir,response_file),encoding='utf-8')
tokenizer,model = get_tokenizer_model(model_name,MODEL_PATHS[model_name],model_cache_dir)
pb = tqdm(questions_df.iterrows(),total=len(questions_df))
right = 0
for i,row in pb:
qid = row['MCQID']
pb.set_description(qid)
if isinstance(already,pd.DataFrame):
if qid in set(already['MCQID']):
continue
country = row['country']
prompt = row['prompt']
print(prompt)
full_res = get_model_response(model_name,prompt,model,tokenizer,temperature,top_p,gpt_azure)
print(full_res)
json_res = get_json_str(full_res)
if isinstance(json_res,dict) and 'answer_choice' in json_res:
try:
final_ans = re.findall(r'[A-Z]',str(json_res['answer_choice']))[0]
if final_ans+'.' not in prompt:
for k,v in json.loads(row['choices']).items():
if v == json_res['answer_choice']:
final_ans = str(k)
break
else:
final_ans = full_res
except:
for k,v in json.loads(row['choices']).items():
if v == json_res['answer_choice']:
final_ans = str(k)
break
else:
final_ans = full_res
else:
try:
final_ans = re.findall(r'[A-Z]',json_res)[0]
except:
final_ans = full_res
write_csv_row(list(row)+[full_res,final_ans],os.path.join(mc_dir,response_file))
if final_ans == row['answer_idx']:
right += 1
pb.set_postfix({'score':right/(i+1)})
def multiple_choice_score(model,mc_dir,mrf,mc_res_file,eval_res_file,wrong_country_ratio_file,country):
df = pd.read_csv(os.path.join(mc_dir,mrf),encoding='utf-8')
df = df[df['country'] == country]
scores = []
for i,row in tqdm(df.iterrows(),total=len(df)):
if str(row['answer_idx']) == str(row['final_ans']):
scores.append(1)
else:
scores.append(0)
df['score'] = scores
final_score = df['score'].mean()
return final_score
if __name__ == "__main__":
parser = argparse.ArgumentParser(description='Choose your model(s) & language(s)')
parser.add_argument('--model',type=str,
help='Provide the model you want to use. Check and choose from the key values of the MODEL_PATHS variable. If you want to test on multiple models, provide multiple model names with ", " between each (e.g., "gpt-4-0125-preview, aya-101").')
parser.add_argument('--model_cache_dir',type=str,default='.cache',
help='Provide the directory saving model caches.')
parser.add_argument('--mc_dir',type=str,default='./mc_data',
help='Provide the directory for the data files from the human annotators.')
parser.add_argument('--questions_file',type=str,default='mc_questions_file.csv',
help='Provide the directory for the data files from the human annotators.')
parser.add_argument('--response_file',type=str,default=None,
help='Provide the filename to save LLM responses.')
parser.add_argument('--temperature',type=int,default=0,
help='Provide generation temperature for LLMs.')
parser.add_argument('--top_p',type=float,default=1,
help='Provide generation top_p for LLMs.')
parser.add_argument("--gpt_azure", type=str2bool, nargs='?',
const=True, default=True,
help="Whether you are using the AzureOpenAI for GPT-models' response generation.")
args = parser.parse_args()
get_model_mc_response(model_name=args.model,
model_cache_dir=args.model_cache_dir,
mc_dir=args.mc_dir,
questions_file=args.questions_file,
response_file=args.response_file,
temperature=args.temperature,
top_p=args.top_p,
gpt_azure=args.gpt_azure)