[Feature] Add multi-prompt generation demo (#568)

* [Feature] Add multi-prompt generation demo

* [Fix] change form in winogrande_gen_XXX.py

* [Fix] make multi prompt demo more directly

* [Fix] fix bug

* [Fix] minor fix

---------

Co-authored-by: yingfhu <yingfhu@gmail.com>

configs/datasets/winogrande/winogrande_gen_a027b6.py

@@ -0,0 +1,49 @@
+from opencompass.openicl.icl_prompt_template import PromptTemplate
+from opencompass.openicl.icl_retriever import ZeroRetriever
+from opencompass.openicl.icl_inferencer import GenInferencer
+from opencompass.openicl.icl_evaluator import AccEvaluator
+from opencompass.datasets import winograndeDataset_V2
+from opencompass.utils.text_postprocessors import first_option_postprocess
+
+winogrande_reader_cfg = dict(
+    input_columns=["opt1", "opt2"],
+    output_column="answer",
+)
+
+winogrande_eval_cfg = dict(
+    evaluator=dict(type=AccEvaluator),
+    pred_role="BOT",
+    pred_postprocessor=dict(type=first_option_postprocess, options='AB'),
+)
+
+_winogrande_prompt = dict(
+    prompt_1="Which of the following is a good sentence:\nA. {opt1}\nB. {opt2}\nAnswer:",
+    prompt_2="Which is a good sentence out of the following:\nA. {opt1}\nB. {opt2}\nAnswer:",
+    prompt_3="Can you identify a good sentence from the following:\nA. {opt1}\nB. {opt2}\nAnswer:",
+)
+
+winogrande_datasets = []
+for _choice in _winogrande_prompt:
+    winogrande_datasets.append(
+        dict(
+            abbr='winogrande_'+_choice,
+            type=winograndeDataset_V2,
+            path="./data/winogrande",
+            reader_cfg=winogrande_reader_cfg,
+            infer_cfg=dict(
+                prompt_template=dict(
+                    type=PromptTemplate,
+                    template=dict(round=[
+                        dict(
+                            role="HUMAN",
+                            prompt=_winogrande_prompt[_choice]
+                        ),
+                    ]),
+                ),
+                retriever=dict(type=ZeroRetriever),
+                inferencer=dict(type=GenInferencer),
+            ),
+            eval_cfg=winogrande_eval_cfg),
+    )
+
+del _choice

configs/eval_multi_prompt_demo.py

@@ -0,0 +1,48 @@
+from mmengine.config import read_base
+from opencompass.models import HuggingFaceCausalLM
+
+
+with read_base():
+    from .datasets.winogrande.winogrande_gen_a027b6 import winogrande_datasets
+
+datasets = [*winogrande_datasets]
+
+_meta_template = dict(
+    round=[
+        dict(role='HUMAN', begin='<|User|>:', end='<eoh>\n'),
+        dict(role='BOT', begin='<|Bot|>:', end='<eoa>\n', generate=True),
+    ],
+)
+
+models=[
+    dict(
+        type=HuggingFaceCausalLM,
+        abbr='internlm-chat-7b-hf',
+        path="internlm/internlm-chat-7b",
+        tokenizer_path='internlm/internlm-chat-7b',
+        tokenizer_kwargs=dict(
+            padding_side='left',
+            truncation_side='left',
+            use_fast=False,
+            trust_remote_code=True,
+        ),
+        max_out_len=100,
+        max_seq_len=2048,
+        batch_size=8,
+        meta_template=_meta_template,
+        model_kwargs=dict(
+            trust_remote_code=True,
+            device_map='auto',
+        ),
+        run_cfg=dict(num_gpus=1, num_procs=1),
+    )
+]
+
+_winogrande_all = [d['abbr'] for d in winogrande_datasets]
+
+summarizer = dict(
+    summary_groups=[
+        {'name': 'winogrande', 'subsets': _winogrande_all},
+        {'name': 'winogrande_std', 'subsets': _winogrande_all, 'std': True},
+    ]
+)

opencompass/summarizers/default.py

@@ -1,6 +1,7 @@
 # flake8: noqa
 # yapf: disable
 import getpass
+import math
 import os.path as osp
 from datetime import datetime
 from typing import List, Optional
@@ -127,6 +128,12 @@ class DefaultSummarizer:
                         results[dataset_abbr] = parsed_results[model_abbr][dataset_abbr][0]
                         eval_modes.append(dataset_eval_mode.get(dataset_abbr, 'unknown'))
                 if len(results) == len(sg['subsets']):
+                    if 'std' in sg and sg['std'] == True:
+                        avg = sum(results[k] for k in results) / len(results)
+                        variance = sum((results[k] - avg)**2 for k in results) / len(results)
+                        metric = 'standard_deviation'
+                        results[metric] = math.sqrt(variance)
+                    else:
                         if 'weights' in sg:
                             numerator = sum(results[k] * sg['weights'][k] for k in sg['weights'])
                             denominator = sum(sg['weights'].values())
@@ -136,12 +143,13 @@ class DefaultSummarizer:
                             denominator = len(results)
                             metric = 'naive_average'
                         results[metric] = numerator / denominator
+
                     eval_modes = list(set(eval_modes))
                     eval_mode = eval_modes[0] if len(eval_modes) == 1 else 'mixed'
-
                     # add to global results
                     raw_results[model_abbr][sg['name']] = results
-                    parsed_results[model_abbr][sg['name']] = [numerator / denominator]
+                    parsed_results[model_abbr][sg['name']] = [results[metric]]
+
                     dataset_metrics[sg['name']] = [metric]
                     dataset_eval_mode[sg['name']] = eval_mode
                 elif len(results) == 0: