From 52eccc4f0efd3ca6f272ae19efb2d7f6cc9c9dec Mon Sep 17 00:00:00 2001 From: klein Date: Tue, 30 Jul 2024 17:51:24 +0800 Subject: [PATCH] [Fix] Fix version mismatch of CIBench (#1380) * update crb * update crbbench * update crbbench * update crbbench * minor update wildbench * [Fix] Update doc of wildbench, and merge wildbench into subjective * [Fix] Update doc of wildbench, and merge wildbench into subjective, fix crbbench * Update crb.md * Update crb_pair_judge.py * Update crb_single_judge.py * Update subjective_evaluation.md * Update openai_api.py * [Update] update wildbench readme * [Update] update wildbench readme * [Update] update wildbench readme, remove crb * Delete configs/eval_subjective_wildbench_pair.py * Delete configs/eval_subjective_wildbench_single.py * Update __init__.py * [Fix] fix version mismatch for CIBench * [Fix] fix version mismatch for CIBench, local runer * [Fix] fix version mismatch for CIBench, local runer, remove oracle mode --------- Co-authored-by: bittersweet1999 <148421775+bittersweet1999@users.noreply.github.com> --- configs/eval_chat_cibench_api.py | 96 ------------ configs/eval_cibench.py | 140 +++++++++++++----- ...al_chat_cibench.py => eval_cibench_api.py} | 71 ++++++--- 3 files changed, 149 insertions(+), 158 deletions(-) delete mode 100644 configs/eval_chat_cibench_api.py rename configs/{eval_chat_cibench.py => eval_cibench_api.py} (60%) diff --git a/configs/eval_chat_cibench_api.py b/configs/eval_chat_cibench_api.py deleted file mode 100644 index 2a80ca87..00000000 --- a/configs/eval_chat_cibench_api.py +++ /dev/null @@ -1,96 +0,0 @@ -from mmengine.config import read_base - -from opencompass.lagent.actions.ipython_interpreter import IPythonInterpreter -from opencompass.lagent.agents.react import CIReAct, ReActProtocol -from opencompass.models.lagent import CodeAgent -from opencompass.models.openai_api import OpenAI -from opencompass.partitioners import SizePartitioner -from opencompass.runners import LocalRunner -from opencompass.tasks import OpenICLInferTask - -with read_base(): - from .datasets.CIBench.CIBench_template_gen_e6b12a import \ - cibench_datasets as datasets - -FORCE_STOP_PROMPT_EN = """You should directly give results based on history information.""" - -FEWSHOT_INSTRUCTION = """\ -You are an assistant who can utilize external tools. -{tool_description} -To use a tool, please response with the following format: -``` -{thought} Think what you need to solve, do you need to use tools? -{action} The tool name, should be one of [{action_names}]. -{action_input} The input to the tool that you want to use. -``` -The tool will give you response after your response using the following format: -``` -{response} the results after call the tool. -``` -Therefore DO NOT generate tool response by yourself. - -Also please follow the guidelines: -1. Always use code interpreter to solve the problem. -2. The generated codes should always in a markdown code block format. -3. The generated codes will be executed in an ipython manner and the results will be cached. -4. Your responded code should always be simple and only solves the problem in current step. - -For example: - -File url: `xxxx` -### Step 1. Load the dataset from the url into a pandas DataFrame named `df`. - -{thought} We should use `pandas` to solve this step. -{action} IPythonInterpreter -{action_input} ```python -import pandas as pd -url = "xxxx" -data = pd.read_csv(url) -``` -{response} The code is succeed without any outputs. - -Let us begin from here! -""" - -IPYTHON_INTERPRETER_DESCRIPTION = '''\ -It can run Python code in a manner as jupyter notebook. The code must be a valid code that contains only python method.''' - -models = [ - dict( - abbr='gpt-3.5-code', - type=CodeAgent, - agent_type=CIReAct, - max_turn=3, - llm=dict( - type=OpenAI, - path='gpt-3.5-turbo', - key='ENV', - query_per_second=1, - max_seq_len=4096, - ), - actions=[ - dict(type=IPythonInterpreter, - description=IPYTHON_INTERPRETER_DESCRIPTION, - user_data_dir='./data/cibench_dataset/datasources') - ], - protocol=dict( - type=ReActProtocol, - call_protocol=FEWSHOT_INSTRUCTION, - force_stop=FORCE_STOP_PROMPT_EN, - finish=dict(role='FINISH', begin='Final Answer:', end='\n'), - ), - batch_size=1, - use_system_role=False, # use `user` role instead of system role - first_system_role=False, # use `user` role of the first instruction prompt - merge_adjacent_role=True, # merge adjacent same user content - ), -] - - -infer = dict( - partitioner=dict(type=SizePartitioner, max_task_size=1000), - runner=dict( - type=LocalRunner, - max_num_workers=16, - task=dict(type=OpenICLInferTask)), -) diff --git a/configs/eval_cibench.py b/configs/eval_cibench.py index 9f05ea39..f2b9a11b 100644 --- a/configs/eval_cibench.py +++ b/configs/eval_cibench.py @@ -1,15 +1,37 @@ +from copy import deepcopy from mmengine.config import read_base -from opencompass.partitioners import SizePartitioner -from opencompass.runners import LocalRunner, SlurmRunner -from opencompass.tasks import OpenICLInferTask -from opencompass.models import OpenAI +from opencompass.models.lagent import LagentAgent +from lagent import ReAct +from lagent.agents.react import ReActProtocol +from opencompass.models.lagent import CodeAgent +from opencompass.lagent.actions.python_interpreter import PythonInterpreter from opencompass.lagent.actions.ipython_interpreter import IPythonInterpreter from opencompass.lagent.agents.react import CIReAct -from opencompass.models.lagent import CodeAgent -from lagent.agents.react import ReActProtocol +from opencompass.models import HuggingFaceCausalLM +from opencompass.partitioners import SizePartitioner +from opencompass.runners import LocalRunner +from opencompass.runners import SlurmSequentialRunner +from opencompass.tasks import OpenICLInferTask +from opencompass.partitioners import NaivePartitioner with read_base(): - from .datasets.CIBench.CIBench_gen_eb42f9 import cibench_datasets as datasets + # Note that it might occur cuda OOM error for hf model + from .models.hf_llama.lmdeploy_llama3_8b_instruct import models as lmdeploy_llama3_8b_instruct_model + + from .summarizers.cibench import summarizer + from .datasets.CIBench.CIBench_template_gen_e6b12a import cibench_datasets as cibench_datasets_template + from .datasets.CIBench.CIBench_generation_gen_8ab0dc import cibench_datasets as cibench_datasets_generation + # Oracle mode for analysis + # from .datasets.CIBench.CIBench_template_oracle_gen_fecda1 import cibench_datasets as cibench_datasets_template_oracle + # from .datasets.CIBench.CIBench_generation_oracle_gen_c4a7c1 import cibench_datasets as cibench_datasets_generation_oracle + +datasets = [] +datasets += cibench_datasets_template +datasets += cibench_datasets_generation +# datasets += cibench_datasets_template_oracle +# datasets += cibench_datasets_generation_oracle + +_origin_models = sum([v for k, v in locals().items() if k.endswith('_model')], []) FORCE_STOP_PROMPT_EN = """You should directly give results based on history information.""" @@ -34,47 +56,87 @@ Also please follow the guidelines: 3. The generated codes will be executed in an ipython manner and the results will be cached. 4. Your responded code should always be simple and only solves the problem in current step. -Begin! +For example: + +File url: `xxxx` +### Step 1. Load the dataset from the url into a pandas DataFrame named `df`. + +{thought} We should use `pandas` to solve this step. +{action} IPythonInterpreter +{action_input} ```python +import pandas as pd +url = "xxxx" +data = pd.read_csv(url) +``` +{response} The code is succeed without any outputs. + +Let us begin from here! """ -models = [ - dict( - abbr='gpt-3.5-turbo', - type=CodeAgent, - agent_type=CIReAct, - mutli_rounds=True, - max_turn=3, - llm=dict( - type=OpenAI, - path='gpt-3.5-turbo', - key='ENV', - query_per_second=1, - max_seq_len=4096, - ), - actions=[ - dict( - type=IPythonInterpreter, - description= - '''It can run Python code in a manner as jupyter notebook. The code must be a valid code that contains only python method. -'''), - ], - protocol=dict( +IPYTHON_INTERPRETER_DESCRIPTION = '''\ +It can run Python code in a manner as jupyter notebook. The code must be a valid code that contains only python method.''' + + + +actions=[dict(type=IPythonInterpreter, user_data_dir='./data/cibench_dataset/datasources', + description=IPYTHON_INTERPRETER_DESCRIPTION)] +protocol=dict( type=ReActProtocol, call_protocol=FEWSHOT_INSTRUCTION, force_stop=FORCE_STOP_PROMPT_EN, - action=dict(role='ACTION', begin='Tool:', end='\n'), - action_input=dict(role='ARGS', begin='Tool Input:', end='\n'), - response=dict(role='RESPONSE', begin='Tool Response:', end='\n'), finish=dict(role='FINISH', begin='Final Answer:', end='\n'), - ), - batch_size=8, - ), -] + ) +work_dir = './outputs/cibench/' + +_agent_models = [] +for m in _origin_models: + m = deepcopy(m) + if 'meta_template' in m and 'round' in m['meta_template']: + round = m['meta_template']['round'] + if all(r['role'].upper() != 'SYSTEM' for r in round): # no system round + if not any('api_role' in r for r in round): + m['meta_template']['round'].append(dict(role='system', begin='System response:', end='\n')) + else: + m['meta_template']['round'].append(dict(role='system', api_role='SYSTEM')) + print(f'WARNING: adding SYSTEM round in meta_template for {m.get("abbr", None)}') + _agent_models.append(m) + +protocol=dict( + type=ReActProtocol, + call_protocol=FEWSHOT_INSTRUCTION, + force_stop=FORCE_STOP_PROMPT_EN, + finish=dict(role='FINISH', begin='Final Answer:', end='\n'), +) + +models = [] +for m in _agent_models: + m = deepcopy(m) + origin_abbr = m.pop('abbr') + abbr = origin_abbr + m.pop('batch_size', None) + m.pop('max_out_len', None) + m.pop('max_seq_len', None) + run_cfg = m.pop('run_cfg', {}) + + agent_model = dict( + abbr=abbr, + summarizer_abbr=origin_abbr, + type=CodeAgent, + agent_type=CIReAct, + max_turn=3, + llm=m, + actions=[dict(type=IPythonInterpreter, user_data_dir='./data/cibench_dataset/datasources', description=IPYTHON_INTERPRETER_DESCRIPTION)], + protocol=protocol, + batch_size=1, + run_cfg=run_cfg, + ) + models.append(agent_model) infer = dict( - partitioner=dict(type=SizePartitioner, max_task_size=50, gen_task_coef=1), + partitioner=dict(type=NaivePartitioner), runner=dict( - type=SlurmRunner, max_num_workers=8, retry=2, + type=LocalRunner, + max_num_workers=4, task=dict(type=OpenICLInferTask)), ) diff --git a/configs/eval_chat_cibench.py b/configs/eval_cibench_api.py similarity index 60% rename from configs/eval_chat_cibench.py rename to configs/eval_cibench_api.py index 2c9b0d54..3c063a03 100644 --- a/configs/eval_chat_cibench.py +++ b/configs/eval_cibench_api.py @@ -1,17 +1,28 @@ from lagent.agents.react import ReActProtocol from mmengine.config import read_base +from opencompass.partitioners import NaivePartitioner -from opencompass.lagent.actions.ipython_interpreter import IPythonInterpreter -from opencompass.lagent.agents.react import CIReAct from opencompass.models.lagent import CodeAgent -from opencompass.models.openai_api import OpenAI -from opencompass.partitioners import SizePartitioner +from opencompass.models import OpenAI from opencompass.runners import LocalRunner from opencompass.tasks import OpenICLInferTask +from opencompass.lagent.actions.ipython_interpreter import IPythonInterpreter +from opencompass.lagent.agents.react import CIReAct with read_base(): - from .datasets.CIBench.CIBench_gen_8ab0dc import \ - cibench_datasets as datasets + from .datasets.CIBench.CIBench_template_gen_e6b12a import cibench_datasets as cibench_datasets_template + from .datasets.CIBench.CIBench_generation_gen_8ab0dc import cibench_datasets as cibench_datasets_generation + # Oracle mode for analysis + # from .datasets.CIBench.CIBench_template_oracle_gen_fecda1 import cibench_datasets as cibench_datasets_template_oracle + # from .datasets.CIBench.CIBench_generation_oracle_gen_c4a7c1 import cibench_datasets as cibench_datasets_generation_oracle + + from .summarizers.cibench import summarizer + +datasets = [] +datasets += cibench_datasets_template +datasets += cibench_datasets_generation +# datasets += cibench_datasets_template_oracle +# datasets += cibench_datasets_generation_oracle FORCE_STOP_PROMPT_EN = """You should directly give results based on history information.""" @@ -56,38 +67,52 @@ Let us begin from here! IPYTHON_INTERPRETER_DESCRIPTION = '''\ It can run Python code in a manner as jupyter notebook. The code must be a valid code that contains only python method.''' + +api_meta_template = dict( + round=[ + dict(role='HUMAN', api_role='HUMAN'), + dict(role='BOT', api_role='BOT', generate=True), + dict(role='SYSTEM', api_role='SYSTEM'), + ], +) + +actions=[dict(type=IPythonInterpreter, user_data_dir='./data/cibench_dataset/datasources', + description=IPYTHON_INTERPRETER_DESCRIPTION)] +protocol=dict( + type=ReActProtocol, + call_protocol=FEWSHOT_INSTRUCTION, + force_stop=FORCE_STOP_PROMPT_EN, + finish=dict(role='FINISH', begin='Final Answer:', end='\n'), + ) + + +work_dir = 'outputs/cibench/' models = [ dict( - abbr='gpt-3.5-code', + abbr='gpt-4o', type=CodeAgent, agent_type=CIReAct, max_turn=3, llm=dict( type=OpenAI, - path='gpt-3.5-turbo', - key='ENV', + path='gpt-4o', + rpm_verbose=True, + retry=99, + meta_template=api_meta_template, query_per_second=1, - max_seq_len=4096, - ), - actions=[ - dict(type=IPythonInterpreter, - description=IPYTHON_INTERPRETER_DESCRIPTION) - ], - protocol=dict( - type=ReActProtocol, - call_protocol=FEWSHOT_INSTRUCTION, - force_stop=FORCE_STOP_PROMPT_EN, - finish=dict(role='FINISH', begin='Final Answer:', end='\n'), + max_seq_len=2048, + temperature=0, ), + actions=actions, + protocol=protocol, batch_size=1, ), ] - infer = dict( - partitioner=dict(type=SizePartitioner, max_task_size=1000), + partitioner=dict(type=NaivePartitioner), runner=dict( type=LocalRunner, - max_num_workers=16, + max_num_workers=4, task=dict(type=OpenICLInferTask)), )