Add ChatGLM C-Eval Evaluator (#10095)

* Add ChatGLM ceval evaluator

* Modify ChatGLM Evaluator Reference

python/llm/dev/benchmark/ceval/eval.py

@@ -24,6 +24,7 @@ from tqdm import tqdm
 from bigdl.llm.utils.common.log4Error import invalidInputError
 from evaluators.qwen import QwenEvaluator
 from evaluators.llama import LlamaEvaluator
+from evaluators.chatglm import ChatGLMEvaluator
 
 
 TASK_NAME_MAPPING = {
@@ -280,7 +281,6 @@ def main(args, evaluator):
 
 if __name__ == "__main__":
     parser = argparse.ArgumentParser()
-    parser.add_argument("--model_family", type=str, default="llama")
     parser.add_argument("--model_path", type=str, default="meta-llama/Llama-2-7b-chat-hf")
     parser.add_argument("--eval_type", type=str, default="validation")
     parser.add_argument("--device", type=str, default="xpu")
@@ -289,22 +289,39 @@ if __name__ == "__main__":
 
     args = parser.parse_args()
 
-    if args.model_family == "llama":
+    # decide the model family
+    model_families = ['llama', 'qwen', 'chatglm']
+
+    model_family = None
+    for family in model_families:
+        if family in args.model_path.lower():
+            model_family = family
+
+    assert model_family is not None, f"Model {args.model_path}'s model family is not implemented"
+
+    if model_family == "llama":
         evaluator = LlamaEvaluator(
             choices=choices,
             model_path=args.model_path,
             device=args.device,
             qtype=args.qtype
         )
-    elif args.model_family == "qwen":
+    elif model_family == "qwen":
         evaluator = QwenEvaluator(
             choices=choices,
             model_path=args.model_path,
             device=args.device,
             qtype=args.qtype
         )
+    elif model_family == "chatglm":
+        evaluator = ChatGLMEvaluator(
+            choices=choices,
+            model_path=args.model_path,
+            device=args.device,
+            qtype=args.qtype
+        )
     else:
         invalidInputError(
             False,
-            "Invalid model_family, currently support llama and qwen only.")
+            "Invalid model_family, currently support llama, qwen, and chatglm only.")
     main(args, evaluator=evaluator)

python/llm/dev/benchmark/ceval/evaluators/chatglm.py

@@ -0,0 +1,211 @@
+#
+# Copyright 2016 The BigDL Authors.
+#
+# 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.
+#
+# refer to https://github.com/THUDM/ChatGLM2-6B/blob/main/evaluation/evaluate_ceval.py
+
+import re
+import torch
+from tqdm import tqdm
+from thefuzz import process
+from transformers import AutoTokenizer
+
+from evaluators.evaluator import Evaluator
+from bigdl.llm.transformers import AutoModelForCausalLM
+from transformers.generation.utils import LogitsProcessorList
+from transformers.generation.logits_process import LogitsProcessor
+
+
+class InvalidScoreLogitsProcessor(LogitsProcessor):
+    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
+        if torch.isnan(scores).any() or torch.isinf(scores).any():
+            scores.zero_()
+            scores[..., 5] = 5e4
+        return scores
+
+
+class ChatGLMEvaluator(Evaluator):
+    def __init__(self, choices, model_path="THUDM/chatglm-6b", device="xpu", qtype="sym_int4"):
+        super(ChatGLMEvaluator, self).__init__(choices, model_path, device, qtype)
+        self.tokenizer = AutoTokenizer.from_pretrained(
+            self.model_path,
+            trust_remote_code=True
+        )
+        self.model = AutoModelForCausalLM.from_pretrained(
+            self.model_path,
+            load_in_low_bit=self.qtype,
+            optimize_model=True,
+            use_cache=True,
+            trust_remote_code=True
+        ).eval().to(self.device)
+  
+    def generate_few_shot_prompt(self, subject, dev_df, cot=False):
+        message = []
+        k = self.k
+        if self.k == -1:
+            k = dev_df.shape[0]
+        message.append(self.format_example(dev_df.iloc[0, :], cot=cot, add_prompt=f"以下是中国关于{subject}考试的单项选择题，请选出其中的正确答案。\n\n"))
+        for i in range(1, k):
+            message.append(self.format_example(dev_df.iloc[i, :], cot=cot))
+        return message
+        
+    def format_example(self, line, include_answer=True, cot=False, add_prompt=''):
+        example = add_prompt + line['question']
+        # print(example)
+        for choice in self.choices:
+            example += f'\n{choice}. {line[f"{choice}"]}'
+        example += '\n答案：'
+        if include_answer:
+            if cot:
+                ans = "让我们一步一步思考，\n" + line["explanation"] + f"\n所以答案是{line['answer']}。"
+            else:
+                ans = line["answer"]
+            m = (example, ans)
+            return m
+        return example
+    
+    def extract_cot_answer(self, line, gen_ans):
+        m = re.findall(r'所以答案是(.+?)。', gen_ans, re.M)
+        if len(m) > 0 and m[-1] in self.choices:
+            return m[-1], True
+        answer_patterns = [
+            r'([ABCD])是正确的',
+            r'选项([ABCD])正确',
+            r'答案为([ABCD])',
+            r'答案是([ABCD])',
+            r'答案([ABCD])',
+            r'选择([ABCD])',
+            r'答案：([ABCD])',
+            r'选择答案([ABCD])'
+        ]
+        # RE extraction
+        for answer_pattern in answer_patterns:
+            m = re.search(answer_pattern, gen_ans, re.M)
+            if m:
+                answer = m.group(1)
+                return answer, False
+        # only containing one choice-character
+        m = re.findall(r'[ABCD]', gen_ans, re.M)
+        if len(m) == 1:
+            answer = m[0]
+            return answer, False
+        answer_word_counter = 0
+        # only containing one choice-context
+        for c in self.choices:
+            if str(line[f'{c}']) in gen_ans:
+                answer = c
+                answer_word_counter += 1
+        if answer_word_counter == 1:
+            return answer, False
+        return '-', False
+    
+    def build_prompt(self, text):
+        return "[Round {}]\n\n问：{}\n\n答：".format(1, text)
+
+    def generate_dist(self, model, tokenizer, query, history, max_length=2048,
+                      do_sample=False, logits_processor=None):
+        
+        if history is None:
+            history = []
+        if logits_processor is None:
+            logits_processor = LogitsProcessorList()
+        logits_processor.append(InvalidScoreLogitsProcessor())
+
+        if not history:
+            prompt = query
+        else:
+            prompt = ""
+            for i, (old_query, response) in enumerate(history):
+                prompt += "[Round {}]\n问：{}\n答：{}\n".format(i, old_query, response)
+            prompt += "[Round {}]\n问：{}\n答：".format(len(history), query)
+
+        # first round prompt
+        inputs = tokenizer([prompt], padding=True, return_tensors="pt",
+                           truncation=True, max_length=max_length).to(model.device)
+        
+        # first round generation
+        outputs = model.generate(**inputs, do_sample=do_sample, max_new_tokens=512)
+
+        # organize intermediate_outputs
+        intermediate_outputs = []
+        for idx in range(len(outputs)):
+            output = outputs.tolist()[idx][len(inputs["input_ids"][idx]):]
+            response = tokenizer.decode(output)
+            intermediate_outputs.append(response)
+
+        # prepare second round prompt
+        extraction_prompt = '综上所述，ABCD中正确的选项是：'
+        answer_texts = [query + intermediate + "\n" + extraction_prompt for intermediate in intermediate_outputs]
+        input_tokens = [self.build_prompt(answer_text) for answer_text in answer_texts]
+        inputs = tokenizer(input_tokens, padding=True, return_tensors="pt", truncation=True, max_length=2048).to(model.device)
+
+        # second round generation
+        outputs = model(**inputs, return_last_logit=True)
+
+        logits = outputs.logits[:, -1]
+        choice_tokens = [tokenizer.encode(choice, add_special_tokens=False)[0] for choice in self.choices]
+        logits = logits[:, choice_tokens]
+        preds = logits.argmax(dim=-1)
+
+        return self.choices[preds]
+
+    @torch.no_grad()
+    def eval_subject(
+        self,
+        subject_name,
+        test_df,
+        eval_type="validation", # "test","validation",
+        dev_df=None,
+        few_shot=False,
+        cot=False,
+    ):
+        if eval_type == "validation":
+            correct_num = 0
+
+            if few_shot:
+                history = self.generate_few_shot_prompt(subject_name, dev_df, cot=cot)
+            else:
+                history = []
+
+            answers = list(test_df['answer'])
+
+            for row_index, row in tqdm(test_df.iterrows(), total=len(test_df)):
+                question = self.format_example(row, include_answer=False, cot=cot)
+
+                if few_shot:
+                    response, _ = self.model.chat(self.tokenizer, question, do_sample=False, history=history)
+                    response = response.strip()
+                    # For ChatGLM, we use answer extraction in answer-only mode too.
+                    ans, direct_extract = self.extract_cot_answer(row, response)
+                else:   # zero-shot by extracting answer from distribution
+                    ans = self.generate_dist(self.model, self.tokenizer, question, do_sample=False, max_length=2048, history=history)
+
+                if ans == answers[row_index]:
+                    correct_num += 1
+
+            correct_ratio = 100*correct_num/len(answers)
+
+            return correct_ratio, None
+        elif eval_type == "test":
+            answers = {}
+            for i, row in tqdm(test_df.iterrows(), total=len(test_df)):
+                question = self.format_example(row)
+                response, _ = self.model.chat(
+                    self.tokenizer,
+                    question,
+                    history=None,
+                )
+                pred = self.extract_answer(response, row)
+                answers[str(i)] = pred
+            return None, answers