OpenCompass/opencompass/datasets/subjective/wildbench.py
2024-12-31 11:01:23 +08:00

438 lines
13 KiB
Python

# flake8: noqa: F401, F403
import json
import re
from collections import defaultdict
from datasets import Dataset, DatasetDict
from opencompass.datasets.subjective.compass_arena_subjective_bench import \
get_element_counts
from opencompass.registry import DICT_POSTPROCESSORS, LOAD_DATASET
from opencompass.utils import get_data_path
from ..base import BaseDataset
from .utils import get_judgeanswer_and_reference
score_prompt = """# Instruction
You are an expert evaluator. Your task is to evaluate the quality of \
the responses generated by AI models.
We will provide you with the user query and an AI-generated responses.
You should first read the user query and the conversation history \
carefully for analyzing the task, and then evaluate the quality of \
the responses based on and rules provided below.
# Conversation between User and AI
## History
<|begin_of_history|>
{history}
<|end_of_history|>
## Current User Query
<|begin_of_query|>
{user_query}
<|end_of_query|>
## AI Response
<|begin_of_response|>
{prediction}
<|end_of_response|>
# Evaluation
## Checklist
<|begin_of_checklist|>
{checklist}
<|end_of_checklist|>
Please use this checklist to guide your evaluation, but do \
not limit your assessment to the checklist.
## Rules
You should compare the above response based on your analysis\
of the user queries and the conversation history.
You should first write down your analysis and the checklist \
that you used for the evaluation, and then provide your \
assessment according to the checklist.
The scores are in the range of 1~10, where 1 means the \
response is very poor and 10 means the response is perfect.
Here are more detailed criteria for the scores:
- Score 1~2: The response is very poor and does not make sense at all.
- Score 3~4: The response is poor and does help user solve the problem\
in a meaningful way.
- Score 5~6: The response is fair but has some issues (e.g., factual \
errors, hallucinations, missing key information).
- Score 7~8: The response is good enough but could be improved in some ways.
- Score 9~10: The response is perfect and provides helpful information that\
can help user solve the problem.
## Output Format
First, please output your analysis for the model response, and then summarize\
your assessment to two aspects: "strengths" and "weaknesses"; Finally, please\
write down your rating for the assessment.
Please provide your evaluation results in the following json format by filling\
in the placeholders in []:
```
{
"strengths": "[analysis for the strengths of the response]",
"weaknesses": "[analysis for the weaknesses of the response]",
"score": "[1~10]"
}
```"""
pair_prompt = """# Instruction
You are an expert evaluator. Your task is to evaluate the quality of the \
responses generated by two AI models.
We will provide you with the user query and a pair of AI-generated \
responses (Response A and Response B).
You should first read the user query and the conversation history \
carefully for analyzing the task, and then evaluate the quality of the \
responses based on and rules provided below.
# Conversation between User and AI
## History
<|begin_of_history|>
{history}
<|end_of_history|>
## Current User Query
<|begin_of_query|>
{user_query}
<|end_of_query|>
## Response A
<|begin_of_response_A|>
{prediction}
<|end_of_response_A|>
## Response B
<|begin_of_response_B|>
{prediction2}
<|end_of_response_B|>
# Evaluation
## Checklist
<|begin_of_checklist|>
{checklist}
<|end_of_checklist|>
Please use this checklist to guide your evaluation, but do not limit your \
assessment to the checklist.
## Rules
You should compare the above two responses based on your analysis of the \
user queries and the conversation history.
You should first write down your analysis and the checklist that you used \
for the evaluation, and then provide your assessment according to the \
checklist.
There are five choices to give your final assessment: ["A++", "A+", \
"A=B", "B+", "B++"], which correspond to the following meanings:
- `A++`: Response A is much better than Response B.
- `A+`: Response A is only slightly better than Response B.
- `A=B`: Response A and B are of the same quality. Please use this \
choice sparingly.
- `B+`: Response B is only slightly better than Response A.
- `B++`: Response B is much better than Response A.
## Output Format
First, please output your analysis for each model response, and \
then summarize your assessment to three aspects: "reason A=B", \
"reason A>B", and "reason B>A", and finally make your choice for \
the final assessment.
Please provide your evaluation results in the following json \
format by filling in the placeholders in []:
```
{
"analysis of A": "[analysis of Response A]",
"analysis of B": "[analysis of Response B]",
"reason of A=B": "[where Response A and B perform equally well]",
"reason of A>B": "[where Response A is better than Response B]",
"reason of B>A": "[where Response B is better than Response A]",
"choice": "[A++ or A+ or A=B or B+ or B++]",
}
```
"""
def parse_conversation(conversation):
# parse conversation into chat dialogue
role_dict = {'user': 'HUMAN', 'assistant': 'assistant'}
chat_round = []
history = ''
if len(conversation) > 0:
for x in conversation[:-1]:
if x['role'] == 'user':
history += 'USER: ' + x['content'] + '\n\n'
elif x['role'] == 'assistant':
history += 'ASSISTANT: ' + x['content'] + '\n\n'
chat_round.append({
'role': role_dict[x['role']],
'content': x['content']
})
last_query = conversation[-1]['content']
chat_round.append({
'role': role_dict[conversation[-1]['role']],
'content': conversation[-1]['content'],
})
chat_round.append({'role': 'assistant', 'content': ''})
return chat_round, last_query, history
@LOAD_DATASET.register_module()
class WildBenchDataset(BaseDataset):
def load(self, path: str, K=-1, eval_mode='pair', *args, **kwargs):
path = get_data_path(path, local_mode=True)
dataset = DatasetDict()
raw_data = []
with open(path, 'r', encoding='utf-8') as file:
for line in file:
item = json.loads(line)
chat_round, last_query, history = parse_conversation(
item['turn'])
checklist_mardkdown = ''
for checklist_item in item['checklist']:
checklist_mardkdown += f'- {checklist_item}\n'
if eval_mode == 'single':
prompt = score_prompt
elif eval_mode == 'pair':
prompt = pair_prompt
else:
assert NotImplementedError(
f'Eval mode {eval_mode} not in single or pair.')
prompt = prompt.replace('{history}', history)
prompt = prompt.replace('{user_query}', last_query)
prompt = prompt.replace('{checklist}', checklist_mardkdown)
raw_data.append({
'dialogue': chat_round,
'history': history,
'prompt': prompt,
'judge': {
'other': None,
'primary_tag': item['primary_tag'],
'secondary_tag': item['secondary_tag'],
'question_id': item['session_id'],
},
})
dataset = Dataset.from_list(raw_data)
return dataset
task_group_new = {
'Information seeking': 'Information/Advice seeking',
'Creative Writing': 'Creative Tasks',
'Coding & Debugging': 'Coding & Debugging',
'Reasoning': 'Planning & Reasoning',
'Editing': 'Creative Tasks',
'Math': 'Math & Data Analysis',
'Planning': 'Planning & Reasoning',
'Brainstorming': 'Creative Tasks',
'Role playing': 'Creative Tasks',
'Advice seeking': 'Information/Advice seeking',
'Data Analysis': 'Math & Data Analysis',
'Others': 'Creative Tasks',
}
def post_process_wildbench_pair(judgement: dict):
judgement = judgement['prediction']
pattern = r"\"choice\": \"(.*?)\""
matched_result = re.findall(pattern, judgement)
if matched_result:
return matched_result[0]
else:
return None
def post_process_wildbench_single(judgement: dict):
judgement = judgement['prediction']
pattern = r"\"score\": \"(.*?)\""
matched_result = re.findall(pattern, judgement)
try:
score = float(matched_result[0])
return {'score': score}
except (ValueError, IndexError) as e:
return None
# if matched_result:
# score = float(matched_result[0])
# else:
# return None
# return {'score': score}
@DICT_POSTPROCESSORS.register_module('wildbench')
def wildbench_postprocess(
output: dict,
output_path: str,
) -> dict:
judged_answers, references = get_judgeanswer_and_reference(
result=output,
filename=output_path,
post_process=post_process_wildbench_pair,
)
if 'base_models' in references[0]:
base_models = references[0]['base_models']
else:
base_models = ['HaiKu', 'gpt4-turbo', 'llama-2-70b-chat-hf']
if isinstance(base_models, str):
base_models = [base_models]
win_base_model = defaultdict(float)
win_compare_model = defaultdict(float)
categories = defaultdict(float)
score_mapping = {'A++': 1, 'A+': 0.5, 'A=B': 0, 'B+': -0.5, 'B++': -1}
for judged_answer, reference in zip(judged_answers, references):
if judged_answer not in score_mapping:
continue
flag = 1 if reference['answer1'] in base_models else -1
score_1 = score_mapping[judged_answer] * flag
score_2 = -score_1
tags = [reference['primary_tag']] + reference['secondary_tag']
for tag in tags:
win_base_model[task_group_new[tag]] += score_1
win_compare_model[task_group_new[tag]] += score_2
categories[task_group_new[tag]] += 1
for capability in categories:
win_base_model[capability] = (win_base_model[capability] /
categories[capability] * 100)
win_base_model[capability] = round(win_base_model[capability], 2)
win_compare_model[capability] = (win_compare_model[capability] /
categories[capability] * 100)
win_compare_model[capability] = round(win_compare_model[capability], 2)
# Calculating the mean of the values
average = sum(win_compare_model.values()) / len(win_compare_model)
# Adding the mean to the dictionary at the beginning
win_compare_model['average'] = average
results = win_compare_model
results['details'] = output
return results
@DICT_POSTPROCESSORS.register_module('wildbench_bradleyterry')
def wildbench_bradleyterry_postprocess(
output: dict,
output_path: str,
) -> dict:
judged_answers, references = get_judgeanswer_and_reference(
result=output,
filename=output_path,
post_process=post_process_wildbench_pair,
)
if 'prediction1' not in references[0]:
raise ValueError(
'prediction1 not in references. Set `keep_predictions=True` for LMEvaluator in dataset config and retry.'
)
if 'prediction2' not in references[0]:
raise ValueError(
'prediction2 not in references. Set `keep_predictions=True` for LMEvaluator in dataset config and retry.'
)
score_mapping = {
'A++': 'model_a',
'A+': 'model_a',
'A=B': 'tie',
'B+': 'model_b',
'B++': 'model_b',
}
results = {}
matches = []
for judged_answer, reference in zip(judged_answers, references):
cur_dict = {}
if judged_answer in score_mapping:
cur_dict['winner'] = score_mapping[judged_answer]
else:
# cur_dict["winner"] = (
# "tie" # Count match as tie if judge answer cannot be parsed.
# )
# Skip if judge answer cannot be parsed
print('Judge answer cannot be parsed. Skipping record...')
continue
cur_dict['primary_tag'] = reference['primary_tag']
# Extract first tag from list and set as categorical level.
# Can be used as categorical variable in Bradley-Terry model
cur_dict['secondary_tag'] = (reference['secondary_tag'][0]
if len(reference['secondary_tag']) > 0
else 'Others')
# Keep original secondary tag list for reference
cur_dict['secondary_tags'] = reference['secondary_tag']
cur_dict['model_a'] = reference['answer1']
cur_dict['model_b'] = reference['answer2']
cur_dict['prediction1'] = reference['prediction1']
cur_dict['prediction2'] = reference['prediction2']
matches.append(cur_dict)
### ---------- Add Style Metadata ---------- ###
matches = get_element_counts(
data=matches,
column='prediction1',
suffix='_a',
)
matches = get_element_counts(
data=matches,
column='prediction2',
suffix='_b',
)
results['matches'] = matches
# results["details"] = output
return results