Add Ceval workflow and modify the result printing (#10140)

* Add c-eval workflow and modify running files

* Modify the chatglm evaluator file

* Modify the ceval workflow for triggering test

* Modify the ceval workflow file

* Modify the ceval workflow file

* Modify ceval workflow

* Adjust the ceval dataset download

* Add ceval workflow dependencies

* Modify ceval workflow dataset download

* Add ceval test dependencies

* Add ceval test dependencies

* Correct the result print

.github/workflows/llm-ceval.yml

@@ -0,0 +1,200 @@
+name: LLM C-Eval
+
+# Cancel previous runs in the PR when you push new commits
+concurrency:
+  group: ${{ github.workflow }}-llm-nightly-test-${{ github.event.pull_request.number || github.run_id }}
+  cancel-in-progress: true
+
+# Controls when the action will run.
+on:
+  schedule:
+    - cron: "00 15 * * 5" # GMT time, 15:00 GMT == 23:00 Beijing Time
+  pull_request:
+    branches: [main]
+    paths:
+      - ".github/workflows/llm-ceval.yml"
+  # Allows you to run this workflow manually from the Actions tab
+  workflow_dispatch:
+    inputs:
+      model_name:
+        description: 'Model names, separated by comma and must be quoted.'
+        required: true
+        type: string
+      precision:
+        description: 'Precisions, separated by comma and must be quoted.'
+        required: true
+        type: string
+      runs-on:
+        description: 'Labels to filter the runners, separated by comma and must be quoted.'
+        default: "accuracy"
+        required: false
+        type: string
+
+# A workflow run is made up of one or more jobs that can run sequentially
+jobs:
+  llm-cpp-build:
+    uses: ./.github/workflows/llm-binary-build.yml
+  # Set the testing matrix based on the event (schedule, PR, or manual dispatch)
+  set-matrix:
+    runs-on: ubuntu-latest
+    outputs:
+      model_name: ${{ steps.set-matrix.outputs.model_name }}
+      precision: ${{ steps.set-matrix.outputs.precision }}
+      runner: ${{ steps.set-matrix.outputs.runner }}
+    steps:
+      - name: set-nightly-env        
+        if: ${{github.event_name == 'schedule'}}
+        env:
+          NIGHTLY_MATRIX_MODEL_NAME: '["chatglm2-6b","chinese-llama2-7b", "Qwen-7B-Chat"]'
+          NIGHTLY_MATRIX_PRECISION: '["sym_int4", "fp8_e5m2"]'
+          NIGHTLY_LABELS: '["self-hosted", "llm", "accuracy-nightly"]'
+        run: |
+            echo "model_name=$NIGHTLY_MATRIX_MODEL_NAME" >> $GITHUB_ENV
+            echo "precision=$NIGHTLY_MATRIX_PRECISION" >> $GITHUB_ENV
+            echo "runner=$NIGHTLY_LABELS" >> $GITHUB_ENV
+
+      - name: set-pr-env
+        if: ${{github.event_name == 'pull_request'}}
+        env:
+          PR_MATRIX_MODEL_NAME: '["Qwen-7B-Chat"]'
+          PR_MATRIX_PRECISION: '["sym_int4", "fp8_e5m2"]'
+          PR_LABELS: '["self-hosted", "llm", "temp-arc01"]'
+        run: |
+            echo "model_name=$PR_MATRIX_MODEL_NAME" >> $GITHUB_ENV
+            echo "precision=$PR_MATRIX_PRECISION" >> $GITHUB_ENV
+            echo "runner=$PR_LABELS" >> $GITHUB_ENV
+      - name: set-manual-env
+        if: ${{github.event_name == 'workflow_dispatch'}}
+        env:
+          MANUAL_MATRIX_MODEL_NAME: ${{format('[ {0} ]', inputs.model_name)}}
+          MANUAL_MATRIX_PRECISION: ${{format('[ {0} ]', inputs.precision)}}
+          MANUAL_LABELS: ${{format('["self-hosted", "llm", {0}]', inputs.runs-on)}}
+        run: |
+            echo "model_name=$MANUAL_MATRIX_MODEL_NAME" >> $GITHUB_ENV
+            echo "precision=$MANUAL_MATRIX_PRECISION" >> $GITHUB_ENV
+            echo "runner=$MANUAL_LABELS" >> $GITHUB_ENV
+      - name: set-matrix
+        id: set-matrix
+        run: |
+            echo "model_name=$model_name" >> $GITHUB_OUTPUT
+            echo "precision=$precision" >> $GITHUB_OUTPUT
+            echo "runner=$runner" >> $GITHUB_OUTPUT
+  llm-ceval-evalution:
+    timeout-minutes: 1200
+    needs: [llm-cpp-build, set-matrix]
+    strategy:
+      fail-fast: false
+      matrix:
+        # include:
+        #   python-version: "3.9"
+        #   model_name: "stablelm-3b-4e1t"
+        #   task: "arc"
+        #   precision: "sym_int4" #options: sym_int4, fp4, mixed_fp4, sym_int8, fp8, mixed_fp8
+        python-version: ["3.9"]
+        model_name: ${{ fromJson(needs.set-matrix.outputs.model_name) }}
+        precision: ${{ fromJson(needs.set-matrix.outputs.precision) }}
+        device: [xpu]
+          
+    runs-on: ${{ fromJson(needs.set-matrix.outputs.runner) }}
+    env:
+      ANALYTICS_ZOO_ROOT: ${{ github.workspace }}
+      ORIGIN_DIR: /mnt/disk1/models
+      CEVAL_HF_HOME: /mnt/disk1/ceval_home
+    steps:
+      - uses: actions/checkout@v3
+      - name: Set up Python ${{ matrix.python-version }}
+        uses: actions/setup-python@v4
+        with:
+          python-version: ${{ matrix.python-version }}
+      - name: Install dependencies
+        shell: bash
+        run: |
+          python -m pip install --upgrade pip
+          python -m pip install --upgrade setuptools==58.0.4
+          python -m pip install --upgrade wheel
+
+      - name: Download llm binary
+        uses: ./.github/actions/llm/download-llm-binary
+
+      - name: Run LLM install (all) test
+        uses: ./.github/actions/llm/setup-llm-env
+        with:
+          extra-dependency: "xpu_2.1"
+      
+      - name: Download models and datasets
+        shell: bash
+        run: |
+          echo "MODEL_PATH=${ORIGIN_DIR}/${{ matrix.model_name }}/" >> "$GITHUB_ENV"
+          MODEL_PATH=${ORIGIN_DIR}/${{ matrix.model_name }}/
+          if [ ! -d $CEVAL_HF_HOME ]; then
+            mkdir -p $CEVAL_HF_HOME
+          fi
+          if [ ! -d "$CEVAL_HF_HOME/data" ]; then
+            mkdir -p "$CEVAL_HF_HOME/data"
+          fi
+          if [ -d "$CEVAL_HF_HOME/data/dev" ]; then
+              rm -rf "$CEVAL_HF_HOME/data/dev"
+          fi
+
+          if [ -d "$CEVAL_HF_HOME/data/test" ]; then
+              rm -rf "$CEVAL_HF_HOME/data/test"
+          fi
+
+          if [ -d "$CEVAL_HF_HOME/data/val" ]; then
+              rm -rf "$CEVAL_HF_HOME/data/val"
+          fi
+          wget -r -nH -nc --no-verbose --cut-dirs=1 ${LLM_FTP_URL}/llm/ceval-exam.zip -P "$CEVAL_HF_HOME/data"
+          echo "DATA_PATH=$CEVAL_HF_HOME/data" >> "$GITHUB_ENV"
+          DATA_PATH=$CEVAL_HF_HOME/data
+          unzip -o "$CEVAL_HF_HOME/data/ceval-exam.zip" -d "$CEVAL_HF_HOME/data"
+          wget -r -nH -nc --no-verbose --cut-dirs=1 ${LLM_FTP_URL}/llm/${{ matrix.model_name }} -P ${ORIGIN_DIR}
+          
+      - name: Install Dependencies
+        shell: bash
+        run: |
+          pip install transformers==4.31.0
+          pip install thefuzz
+          pip install tiktoken
+          pip install transformers_stream_generator
+
+      - name: Run C-Eval
+        shell: bash
+        working-directory: ${{ github.workspace }}/python/llm/dev/benchmark/ceval
+        env:
+          USE_XETLA: OFF
+          SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS: 1
+        run: |
+          source /opt/intel/oneapi/setvars.sh
+          python eval.py \
+          --model_path ${MODEL_PATH} \
+          --eval_type validation \
+          --device xpu \
+          --eval_data_path ${DATA_PATH} \
+          --qtype ${{ matrix.precision }}
+        
+      - uses: actions/upload-artifact@v3
+        with:
+          name: ceval_results
+          path:
+            ${{ github.workspace }}/python/llm/dev/benchmark/ceval/results/**
+
+  llm-ceval-summary:
+    if: ${{ always() }}
+    needs: llm-ceval-evalution
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v3
+      - name: Set up Python 3.9
+        uses: actions/setup-python@v4
+        with:
+          python-version: 3.9
+      - name: Download all results
+        uses: actions/download-artifact@v3
+        with:
+          name: ceval_results
+          path: results
+      - name: Summarize the results
+        shell: bash
+        run: |
+          ls results
+          python ${{ github.workspace }}/python/llm/dev/benchmark/ceval/organize_results.py results/

python/llm/dev/benchmark/ceval/README.md

@@ -19,7 +19,6 @@ bash run.sh
 + `run.sh`
 ```shell
 python eval.py \
-    --model_family llama \
     --model_path "path to model" \
     --eval_type validation \
     --device xpu \

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

@@ -222,7 +222,7 @@ hard_list = [
 choices = ["A", "B", "C", "D"]
 
 
-def cal_ceval(res):
+def cal_ceval(res, model_path, qtype):
     acc_sum_dict = dict()
     acc_norm_sum_dict = dict()
     cnt_dict = dict()
@@ -244,13 +244,22 @@ def cal_ceval(res):
             hard_acc_sum += float(res[tt])
         acc_sum_dict[class_] += float(res[tt])
         cnt_dict[class_] += 1
-    print("\n\n\n")
+    
-    for k in ["STEM", "Social Science", "Humanities", "Other"]:
+    result_lst = []
-        if k in cnt_dict:
+    subject_names = ["STEM", "Social Science", "Humanities", "Other", "Hard", "Average"]
-            print("%s acc: %.2f " % (k, acc_sum_dict[k] / cnt_dict[k]))
+    for value in subject_names:
-    if hard_cnt > 0:
+        if value == "Hard":
-        print("Hard acc:%.2f " % (hard_acc_sum / hard_cnt))
+            result_lst.append(f"{hard_acc_sum / hard_cnt:.2f}")
-    print("AVERAGE acc:%.2f " % (acc_sum / cnt))
+        elif value == "Average":
+            result_lst.append(f"{acc_sum / cnt:.2f}")
+        else:
+            result_lst.append(f"{acc_sum_dict[value] / cnt_dict[value]:.2f}")
+    
+    if not os.path.exists('results/'):
+        os.mkdir('results/')
+
+    dump_dict = {"Model Name": model_path.split('/')[-2], "Precision": qtype, "Results": result_lst}
+    json.dump(dump_dict, open(f'results/{dump_dict["Model Name"]}_{dump_dict["Precision"]}.json','w'), ensure_ascii=False, indent=4)
 
 
 def main(args, evaluator):
@@ -262,8 +271,9 @@ def main(args, evaluator):
             )
             val_df = pd.read_csv(val_file_path)
             score, _ = evaluator.eval_subject(subject_name, val_df, args.eval_type)
+            torch.xpu.empty_cache()
             result[subject_name] = score
-        cal_ceval(result)
+        cal_ceval(result, args.model_path, args.qtype)
     elif args.eval_type == "test":
         all_answers = {}
         for subject_name in tqdm(TASK_NAME_MAPPING.keys()):
@@ -272,6 +282,7 @@ def main(args, evaluator):
             )
             test_df = pd.read_csv(test_file_path)
             _, answers = evaluator.eval_subject(subject_name, test_df, args.eval_type)
+            torch.xpu.empty_cache()
             all_answers[subject_name] = answers
         json.dump(all_answers, open('submission.json','w'), ensure_ascii=False, indent=4)
     else:
@@ -297,7 +308,7 @@ if __name__ == "__main__":
         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"
+    assert model_family is not None, f"Model {args.model_path}'s evaluator is not implemented"
 
     if model_family == "llama":
         evaluator = LlamaEvaluator(

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

@@ -60,7 +60,7 @@ class ChatGLMEvaluator(Evaluator):
             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=''):
+    def format_example(self, line, include_answer=False, cot=False, add_prompt=''):
         example = add_prompt + line['question']
         # print(example)
         for choice in self.choices:
@@ -110,6 +110,51 @@ class ChatGLMEvaluator(Evaluator):
             return answer, False
         return '-', False
     
+    def extract_choice(self, gen, prompt, choice_list):
+        res = re.search(
+            r"(?:(?:选|选择|选定)[：:]?\s*|(?:(?:答案|选项)(?![^ABCD]{0,10}?(?:不|非)[^ABCD]{0,10}?(?:是|选|为|：|:|】))[^ABCD]{0,10}?(?:是|选|为|：|:|】))[^ABCD]{0,10}?)(A|B|C|D)(?:选项)?(?:\)|。|\.|，|,|．|、|A|B|C|D|$|：|:|\)|）)",
+            gen,
+        )
+
+        if res is None:
+            res = re.search(
+                r"(A|B|C|D)(?:选?项)?(?![^ABCD]{0,4}?(?:不|非)[^ABCD]{0,4}?(?:正确|对[的，。：]|符合))[^ABCD]{0,4}?(?:正确|对[的，。：]|符合)",
+                gen,
+            )
+
+        if res is None:
+            res = re.search(r"^[\(（]?(A|B|C|D)(?:。|\)|）|\.|，|,|．|：|:|$)", gen)
+
+        if res is None:
+            res = re.search(r"(?<![a-zA-Z])(A|B|C|D)(?![a-zA-Z=])", gen)
+
+        if res is None:
+            return self.choices[choice_list.index(process.extractOne(gen, choice_list)[0])]
+        return res.group(1)
+
+    def process_before_extraction(self, gen, question, choice_dict):
+
+        question_split = question.rstrip("。").split("。")[-1].split("_")
+
+        if len(question_split[0].strip()) > 4:
+            gen = gen.replace(question_split[0], "答案是")
+        if len(question_split[-1].strip()) > 4:
+            gen = gen.replace(question_split[-1], "")
+
+        for key, val in sorted(choice_dict.items(), key=lambda x: len(x[1]), reverse=True):
+            gen = gen.replace(val.rstrip("。"), key)
+        return gen
+
+    def extract_answer(self, response, row):
+        prompt = row["question"]
+        gen = self.process_before_extraction(
+            response, prompt, {choice: row[choice] for choice in self.choices}
+        )
+        if not isinstance(prompt, str):
+            prompt = prompt[0]
+        pred = self.extract_choice(gen, prompt, [row[choice] for choice in self.choices])
+        return pred
+
     def build_prompt(self, text):
         return "[Round {}]\n\n问：{}\n\n答：".format(1, text)
 
@@ -168,7 +213,7 @@ class ChatGLMEvaluator(Evaluator):
         eval_type="validation", # "test","validation",
         dev_df=None,
         few_shot=False,
-        cot=False,
+        cot=True,
     ):
         if eval_type == "validation":
             correct_num = 0
@@ -200,12 +245,7 @@ class ChatGLMEvaluator(Evaluator):
         elif eval_type == "test":
             answers = {}
             for i, row in tqdm(test_df.iterrows(), total=len(test_df)):
-                question = self.format_example(row)
+                question = self.format_example(row, include_answer=False, cot=cot)
-                response, _ = self.model.chat(
+                answers[str(i)] = self.generate_dist(self.model, self.tokenizer, question, do_sample=False, max_length=2048, history=[])
-                    self.tokenizer,
+
-                    question,
-                    history=None,
-                )
-                pred = self.extract_answer(response, row)
-                answers[str(i)] = pred
             return None, answers

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

@@ -0,0 +1,68 @@
+#
+# 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.
+#
+
+import os
+import sys
+import json
+
+if __name__ == '__main__':
+    result_path = sys.argv[1]
+
+    column_size = [25, 15, 10, 18, 15, 10, 10, 10]
+    pad_string = lambda x, l: [i.ljust(j) for i, j in zip(x, l)]
+    column_names = ["Model Name", "Precision", "STEM", "Social Science", "Humanities", "Other", "Hard", "Average"]
+
+    print(f'\nDumping results for C-Eval score:\n')
+    print(' '.join(pad_string(column_names, column_size)))
+    print()
+
+    file_lst = os.listdir(result_path)
+    file_lst = [f'{result_path}/{i}' for i in file_lst]
+
+    organized_dict = {}   # {'Qwen-7B': {'sym_int4': [], 'mixed_fp4': }}
+    for file in file_lst:
+        # Read the JSON file
+        with open(file, 'r') as file:
+            data = json.load(file)
+
+            result_lst = [data['Model Name'], data['Precision']]
+
+            result_lst += data['Results']
+
+            # store in the organized dictionary
+            try:
+                organized_dict[data['Model Name']][data['Precision']] = result_lst
+            except:
+                organized_dict[data['Model Name']] = {}
+                organized_dict[data['Model Name']][data['Precision']] = result_lst
+
+    # define the print precision order    
+    precision_order = ['sym_int4', 'mixed_fp4', 'fp4', 'sym_int8', 'fp8_e4m3', 'fp8_e5m2', 'mixed_fp8']
+
+    # print the results
+    for model_name in organized_dict.keys():
+        for precision in precision_order:
+            try:
+                print(' '.join(pad_string(organized_dict[model_name][precision], column_size)))
+            except KeyError:
+                continue
+
+        # separate between models
+        print()
+
+
+
+