[Fix] Update alignmentbench (open-compass#704)

* update alignmentbench

* update alignmentbench

* update alignmentbench

configs/datasets/subjective_cmp/alignment_bench.py

@@ -1,5 +1,3 @@
-from os import getenv as gv
-
 from opencompass.openicl.icl_prompt_template import PromptTemplate
 from opencompass.openicl.icl_retriever import ZeroRetriever
 from opencompass.openicl.icl_inferencer import GenInferencer
@@ -15,10 +13,10 @@
 subjective_all_sets = [
     "alignment_bench",
 ]
-data_path =gv('WORKDIR')+"data/subjective/alignment_bench"
+data_path ="data/subjective/alignment_bench"
 
-alignment_bench_config_path = gv('WORKDIR')+"data/subjective/alignment_bench/config"
-alignment_bench_config_name = 'multi-dimension'
+alignment_bench_config_path = "data/subjective/alignment_bench/"
+alignment_bench_config_name = 'config/multi-dimension'
 
 subjective_datasets = []
 

configs/alignment_bench.py → configs/eval_subjective_alignbench.py

@@ -11,25 +11,19 @@
 
 datasets = [*subjective_datasets]
 
-from opencompass.models import HuggingFaceCausalLM, HuggingFace, OpenAI, HuggingFaceChatGLM3
+from opencompass.models import HuggingFaceCausalLM, HuggingFace, OpenAIAllesAPIN, HuggingFaceChatGLM3
 from opencompass.partitioners import NaivePartitioner
 from opencompass.partitioners.sub_naive import SubjectiveNaivePartitioner
 from opencompass.runners import LocalRunner
 from opencompass.runners import SlurmSequentialRunner
 from opencompass.tasks import OpenICLInferTask
 from opencompass.tasks.subjective_eval import SubjectiveEvalTask
 from opencompass.summarizers import AlignmentBenchSummarizer
-models = [*hf_baichuan2_7b]#, *hf_chatglm3_6b, *hf_internlm_chat_20b, *hf_qwen_7b_chat, *hf_qwen_14b_chat]
 
-api_meta_template = dict(
-    round=[
-        dict(role='HUMAN', api_role='HUMAN'),
-        dict(role='BOT', api_role='BOT', generate=True)
-    ],
-    reserved_roles=[
-        dict(role='SYSTEM', api_role='SYSTEM'),
-    ],
-)
+
+# -------------Inferen Stage ----------------------------------------
+
+models = [*hf_baichuan2_7b]#, *hf_chatglm3_6b, *hf_internlm_chat_20b, *hf_qwen_7b_chat, *hf_qwen_14b_chat]
 
 infer = dict(
     partitioner=dict(type=NaivePartitioner),
@@ -42,6 +36,10 @@
 )
 
 
+# -------------Evalation Stage ----------------------------------------
+
+
+## ------------- JudgeLLM Configuration
 api_meta_template = dict(
     round=[
         dict(role='HUMAN', api_role='HUMAN'),
@@ -50,44 +48,35 @@
 )
 
 judge_model = dict(
-        type=HuggingFaceChatGLM3,
-        abbr='chatglm3-6b-hf',
-        path='THUDM/chatglm3-6b',
-        tokenizer_path='THUDM/chatglm3-6b',
-        model_kwargs=dict(
-            device_map='auto',
-            trust_remote_code=True,
-        ),
-        tokenizer_kwargs=dict(
-            padding_side='left',
-            truncation_side='left',
-            trust_remote_code=True,
-        ),
+        abbr='GPT4-Turbo',
+        type=OpenAIAllesAPIN, path='gpt-4-1106-preview',
+        key='xxxx',  # The key will be obtained from $OPENAI_API_KEY, but you can write down your key here as well
+        url='xxxx',
         meta_template=api_meta_template,
-        max_out_len=100,
-        max_seq_len=4096,
-        batch_size=1,
-        run_cfg=dict(num_gpus=1, num_procs=1)
-    )
+        query_per_second=16,
+        max_out_len=2048,
+        max_seq_len=2048,
+        batch_size=8
+)
 
+## ------------- Evaluation Configuration
 eval = dict(
     partitioner=dict(
         type=SubjectiveNaivePartitioner,
         mode='singlescore',
         models = [*hf_baichuan2_7b]
     ),
     runner=dict(
-        type=SlurmSequentialRunner,
-        partition='llmeval',
-        quotatype='auto',
-        max_num_workers=256,
+        type=LocalRunner,
+        max_num_workers=2,
         task=dict(
             type=SubjectiveEvalTask,
             judge_cfg=judge_model
         )),
 )
-work_dir = gv('WORKDIR')+'alignment_bench/'
 
 summarizer = dict(
     type=AlignmentBenchSummarizer,
-)
+)
+
+work_dir = 'outputs/alignment_bench/'

configs/models/hf_internlm/hf_internlm_chat_20b.py

@@ -3,7 +3,7 @@
 
 _meta_template = dict(
     round=[
-        dict(role='HUMAN', begin='<|User|>:', end='<eoh>\n'),
+        dict(role='HUMAN', begin='<|User|>:', end='\n'),
         dict(role='BOT', begin='<|Bot|>:', end='<eoa>\n', generate=True),
     ],
 )

configs/models/hf_internlm/hf_internlm_chat_7b.py

@@ -3,7 +3,7 @@
 
 _meta_template = dict(
     round=[
-        dict(role='HUMAN', begin='<|User|>:', end='<eoh>\n'),
+        dict(role='HUMAN', begin='<|User|>:', end='\n'),
         dict(role='BOT', begin='<|Bot|>:', end='<eoa>\n', generate=True),
     ],
 )

configs/models/hf_internlm/hf_internlm_chat_7b_8k.py

@@ -3,7 +3,7 @@
 
 _meta_template = dict(
     round=[
-        dict(role='HUMAN', begin='<|User|>:', end='<eoh>\n'),
+        dict(role='HUMAN', begin='<|User|>:', end='\n'),
         dict(role='BOT', begin='<|Bot|>:', end='<eoa>\n', generate=True),
     ],
 )

opencompass/models/__init__.py

@@ -17,7 +17,7 @@
 from .mixtral import Mixtral  # noqa: F401
 from .modelscope import ModelScope, ModelScopeCausalLM  # noqa: F401, F403
 from .moonshot_api import MoonShot  # noqa: F401
-from .openai_api import OpenAI  # noqa: F401
+from .openai_api import OpenAI, OpenAIAllesAPIN  # noqa: F401
 from .pangu_api import PanGu  # noqa: F401
 from .sensetime_api import SenseTime  # noqa: F401
 from .turbomind import TurboMindModel  # noqa: F401

opencompass/models/openai_api.py

@@ -310,3 +310,121 @@ def bin_trim(self, prompt: str, num_token: int) -> str:
         elif self.mode == 'rear':
             prompt = sep.join(words[:l])
         return prompt
+
+
+@MODELS.register_module(name=['OpenAIAllesAPIN'])
+class OpenAIAllesAPIN(OpenAI):
+    """Model wrapper around OpenAI-AllesAPIN.
+
+    Args:
+        path (str): The name of OpenAI's model.
+        url (str): URL to AllesAPIN.
+        key (str): AllesAPIN key.
+        query_per_second (int): The maximum queries allowed per second
+            between two consecutive calls of the API. Defaults to 1.
+        max_seq_len (int): Unused here.
+        meta_template (Dict, optional): The model's meta prompt
+            template if needed, in case the requirement of injecting or
+            wrapping of any meta instructions.
+        retry (int): Number of retires if the API call fails. Defaults to 2.
+    """
+
+    is_api: bool = True
+
+    def __init__(self,
+                 path: str,
+                 url: str,
+                 key: str,
+                 query_per_second: int = 1,
+                 rpm_verbose: bool = False,
+                 max_seq_len: int = 2048,
+                 meta_template: Optional[Dict] = None,
+                 retry: int = 2):
+        super().__init__(path=path,
+                         max_seq_len=max_seq_len,
+                         query_per_second=query_per_second,
+                         rpm_verbose=rpm_verbose,
+                         meta_template=meta_template,
+                         retry=retry)
+        self.url = url
+        self.headers = {
+            'alles-apin-token': key,
+            'content-type': 'application/json',
+        }
+
+    def _generate(self, input: str or PromptList, max_out_len: int,
+                  temperature: float) -> str:
+        """Generate results given an input.
+
+        Args:
+            inputs (str or PromptList): A string or PromptDict.
+                The PromptDict should be organized in OpenCompass'
+                API format.
+            max_out_len (int): The maximum length of the output.
+            temperature (float): What sampling temperature to use,
+                between 0 and 2. Higher values like 0.8 will make the output
+                more random, while lower values like 0.2 will make it more
+                focused and deterministic.
+
+        Returns:
+            str: The generated string.
+        """
+        assert isinstance(input, (str, PromptList))
+
+        if isinstance(input, str):
+            messages = [{'role': 'user', 'content': input}]
+        else:
+            messages = []
+            for item in input:
+                msg = {'content': item['prompt']}
+                if item['role'] == 'HUMAN':
+                    msg['role'] = 'user'
+                elif item['role'] == 'BOT':
+                    msg['role'] = 'assistant'
+                elif item['role'] == 'SYSTEM':
+                    msg['role'] = 'system'
+                messages.append(msg)
+            # model can be response with user and system
+            # when it comes with agent involved.
+            assert msg['role'] in ['user', 'system']
+        data = {
+            'model': self.path,
+            'messages': messages,
+        }
+
+        for _ in range(self.retry):
+            self.wait()
+            raw_response = requests.post(self.url,
+                                         headers=self.headers,
+                                         data=json.dumps(data))
+            try:
+                response = raw_response.json()
+            except requests.JSONDecodeError:
+                self.logger.error('JsonDecode error, got',
+                                  str(raw_response.content))
+                continue
+            if raw_response.status_code == 200 and response[
+                    'msgCode'] == '10000':
+                data = response['data']
+                choices = data['choices']
+                if choices is None:
+                    self.logger.error(data)
+                else:
+                    return choices[0]['message']['content'].strip()
+            self.logger.error(response['msg'])
+
+        raise RuntimeError('API call failed.')
+
+    def get_token_len(self, prompt: str) -> int:
+        """Get lengths of the tokenized string. Only English and Chinese
+        characters are counted for now. Users are encouraged to override this
+        method if more accurate length is needed.
+
+        Args:
+            prompt (str): Input string.
+
+        Returns:
+            int: Length of the input tokens
+        """
+        enc = self.tiktoken.encoding_for_model(self.path)
+        return len(enc.encode(prompt))

opencompass/runners/local_api.py

@@ -172,7 +172,8 @@ def __init__(self,
         self.max_num_workers = max_num_workers
         self.concurrent_users = concurrent_users
         assert task['type'] in [
-            'OpenICLInferTask', 'opencompass.tasks.OpenICLInferTask'
+            'OpenICLInferTask',
+            'opencompass.tasks.OpenICLInferTask',
         ], 'Only supported for api infer task.'
 
     def launch(self, tasks: List[Dict[str, Any]]) -> List[Tuple[str, int]]:

opencompass/summarizers/alignmentbench.py

@@ -15,7 +15,7 @@
 except ImportError:
     from_csv = None
 
-from opencompass.utils import dataset_abbr_from_cfg
+from opencompass.utils import dataset_abbr_from_cfg, model_abbr_from_cfg
 
 CATEGORIES = {
     '中文推理': ['数学计算', '逻辑推理'],
@@ -91,6 +91,10 @@ class AlignmentBenchSummarizer:
     def __init__(self, config: ConfigDict) -> None:
         self.tasks = []
         self.cfg = config
+        self.eval_model_cfgs = self.cfg['eval']['partitioner']['models']
+        self.eval_model_abbrs = [
+            model_abbr_from_cfg(model) for model in self.eval_model_cfgs
+        ]
 
     def summarize(self,
                   time_str: str = datetime.now().strftime('%Y%m%d_%H%M%S')):
@@ -116,6 +120,8 @@ def summarize(self,
         fout2 = osp.join(output_dir, 'capability.csv')
         fout_flag, fout_flag2 = 0, 0
         for subdir in os.listdir(results_folder):
+            if subdir not in self.eval_model_abbrs:
+                continue
             subdir_path = os.path.join(results_folder, subdir)
             if os.path.isdir(subdir_path):
                 model = subdir