Files
Search-R1/verl/trainer/main_eval.py
PeterGriffinJin 068516be64 Initial commit
2025-02-28 15:16:19 +00:00

70 lines
2.2 KiB
Python

# Copyright 2024 Bytedance Ltd. and/or its affiliates
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
Offline evaluate the performance of a generated file using reward model and ground truth verifier.
The input is a parquet file that contains N generated sequences and (optional) the ground truth.
"""
import hydra
from verl.utils.fs import copy_local_path_from_hdfs
from verl.utils.reward_score import math, gsm8k
import pandas as pd
import numpy as np
def select_reward_fn(data_source):
if data_source == 'lighteval/MATH':
return math.compute_score
else:
raise NotImplementedError
@hydra.main(config_path='config', config_name='evaluation', version_base=None)
def main(config):
local_path = copy_local_path_from_hdfs(config.data.path)
dataset = pd.read_parquet(local_path)
prompts = dataset[config.data.prompt_key]
responses = dataset[config.data.response_key]
data_sources = dataset[config.data.data_source_key]
reward_model_data = dataset[config.data.reward_model_key]
passes = 0
total = len(dataset)
for i in range(total):
response_lst = responses[i]
data_source = data_sources[i]
# select reward score based on data_source
prompt = prompts[i]
reward_data = reward_model_data[i]
reward_fn = select_reward_fn(data_source)
ground_truth = reward_data['ground_truth']
score_lst = []
for r in response_lst:
score = reward_fn(r, ground_truth)
score_lst.append(score)
max_score = np.max(score_lst)
if max_score == 1:
passes += 1
print(f'pass@5: {passes / total}')
if __name__ == '__main__':
main()