Add C-Eval HTML report (#10294)

* Add C-Eval HTML report

* Fix C-Eval workflow pr trigger path

* Fix C-Eval workflow typos

* Add permissions to C-Eval workflow

* Fix C-Eval workflow typo

* Add pandas dependency

* Fix C-Eval workflow typo

.github/workflows/llm-c-evaluation.yml

@@ -15,7 +15,7 @@ on:
   pull_request:
     branches: [main]
     paths:
-      - ".github/workflows/llm-ceval.yml"
+      - ".github/workflows/llm-c-evaluation.yml"
   # Allows you to run this workflow manually from the Actions tab
   workflow_dispatch:
     inputs:
@@ -84,7 +84,8 @@ jobs:
             echo "model_name=$model_name" >> $GITHUB_OUTPUT
             echo "precision=$precision" >> $GITHUB_OUTPUT
             echo "runner=$runner" >> $GITHUB_OUTPUT
-  llm-ceval-evaluation:
+
+  llm-c-evaluation:
     timeout-minutes: 1200
     needs: [llm-cpp-build, set-matrix]
     strategy:
@@ -94,7 +95,7 @@ jobs:
         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 }}
@@ -112,6 +113,11 @@ jobs:
           python -m pip install --upgrade pip
           python -m pip install --upgrade setuptools==58.0.4
           python -m pip install --upgrade wheel
+          pip install einops
+          pip install thefuzz
+          pip install tiktoken
+          pip install transformers==4.31.0
+          pip install transformers_stream_generator
 
       - name: Download llm binary
         uses: ./.github/actions/llm/download-llm-binary
@@ -150,15 +156,6 @@ jobs:
           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 einops
-          pip install thefuzz
-          pip install tiktoken
-          pip install transformers==4.31.0
-          pip install transformers_stream_generator
 
       - name: Run C-Eval
         shell: bash
@@ -183,7 +180,7 @@ jobs:
 
   llm-ceval-summary:
     if: ${{ always() }}
-    needs: llm-ceval-evaluation
+    needs: llm-c-evaluation
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@v3
@@ -191,13 +188,80 @@ jobs:
         uses: actions/setup-python@v4
         with:
           python-version: 3.9
-      - name: Download all results
+      - name: Install dependencies
+        shell: bash
+        run: |
+          pip install --upgrade pip
+          pip install pandas==1.5.3
+
+      - name: Download ceval results
         uses: actions/download-artifact@v3
         with:
           name: ceval_results
           path: results
+
       - name: Summarize the results
         shell: bash
         run: |
           ls results
-          python ${{ github.workspace }}/python/llm/dev/benchmark/ceval/organize_results.py results/
+          echo "DATE=$(date +%Y-%m-%d)" >> $GITHUB_ENV
+          python ${{ github.workspace }}/python/llm/dev/benchmark/ceval/organize_results.py results/ results/
+
+      - name: Set artifact file path
+        run: echo "ARTIFACT_PATH=results/results_${{ env.DATE }}.csv" >> $GITHUB_ENV
+
+      - uses: actions/upload-artifact@v3
+        with:
+          name: results_${{ env.DATE }}
+          path: ${{ env.ARTIFACT_PATH }}
+
+  llm-ceval-html:
+    if: ${{github.event_name == 'schedule' || github.event_name == 'pull_request'}}
+    needs: [llm-c-evaluation, llm-ceval-summary]
+    runs-on: ["self-hosted", "llm", "accuracy1", "accuracy-nightly"]
+    steps:
+      - uses: actions/checkout@f43a0e5ff2bd294095638e18286ca9a3d1956744 # actions/checkout@v3
+      - name: Set up Python 3.9
+        uses: actions/setup-python@v4
+        with:
+          python-version: 3.9
+      - name: Install dependencies
+        shell: bash
+        run: |
+          pip install --upgrade pip
+          pip install numpy
+          pip install pandas==1.5.3
+          pip install jsonlines pytablewriter regex
+      
+      - name: Set output path
+        shell: bash
+        run: |    
+          echo "DATE=$(date +%Y-%m-%d)" >> $GITHUB_ENV
+          if ${{github.event_name == 'pull_request'}}; then
+            echo 'ACC_FOLDER=/home/arda/ceval-action-runners/pr-accuracy-data' >> $GITHUB_ENV
+          fi
+          if ${{github.event_name == 'schedule'}}; then
+            echo 'ACC_FOLDER=/home/arda/ceval-action-runners/nightly-accuracy-data' >> $GITHUB_ENV
+          fi
+
+      - name: Create ceval results directory if not exists
+        run: |
+          if [ ! -d "${{ env.ACC_FOLDER }}" ]; then
+            mkdir -p "${{ env.ACC_FOLDER }}"
+          fi
+
+      - name: Download ceval results
+        uses: actions/download-artifact@v3
+        with:
+          name: results_${{ env.DATE }}
+          path:  ${{ env.ACC_FOLDER }}
+          rename: results_${{ env.DATE }}.csv
+
+      - name: Update HTML
+        working-directory: ${{ github.workspace }}/python/llm/test/benchmark/ceval/
+        shell: bash
+        run: |
+          python ceval_csv_to_html.py -f $ACC_FOLDER
+          if ${{github.event_name == 'schedule'}}; then
+            python update_html_in_parent_folder.py -f $ACC_FOLDER
+          fi

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

@@ -15,11 +15,17 @@
 #
 
 import os
+import pdb
 import sys
+import csv
 import json
+import datetime
+import pandas as pd
+
 
 if __name__ == '__main__':
     result_path = sys.argv[1]
+    output_path = sys.argv[2]
 
     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)]
@@ -49,20 +55,40 @@ if __name__ == '__main__':
                 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']
-
+    # define the print precision order
+    model_order = ['chatglm2-6b', 'chinese-llama2-7b', 'Qwen-7B-Chat']
+    precision_order = ['sym_int4', 'fp8_e5m2']
+    
     # print the results
     for model_name in organized_dict.keys():
         for precision in precision_order:
             try:
+                # print the result
                 print(' '.join(pad_string(organized_dict[model_name][precision], column_size)))
             except KeyError:
-                continue
-
+                pass
         # separate between models
         print()
-
-
+    
+    # initialize the csv file
+    current_date = datetime.datetime.now().strftime("%Y-%m-%d")
+    file_name = f'results_{current_date}.csv'
+    file_name = os.path.join(output_path, file_name) if output_path else file_name
+    print('Writing to', file_name)
+
+    with open(file_name, mode='w', newline='') as csv_file:
+        writer = csv.writer(csv_file)
+        
+        headers = ["Model Name", "Precision", 'STEM', 'Social Science', 'Humanities', 'Other', 'Hard', 'Average']
+        writer.writerow(headers)
+
+        # print the results
+        for model_name in model_order:
+            for precision in precision_order:
+                try:
+                    # write the result to the csv row
+                    writer.writerow(organized_dict[model_name][precision])
+                except KeyError:
+                    writer.writerow([model_name, precision]+[pd.NA for i in range(len(headers[2:]))])
 
 

python/llm/test/benchmark/ceval/ceval_csv_to_html.py

@@ -0,0 +1,138 @@
+#
+# 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.
+#
+
+# Python program to convert CSV to HTML Table
+
+import os
+import sys
+import argparse
+import numpy as np
+import pandas as pd
+
+def highlight_vals(val, max=3.0, color1='red', color2='green', color3='yellow', is_last=False):
+    if isinstance(val, float):
+        if val > max:
+            return 'background-color: %s' % color2
+        elif val <= -max:
+            return 'background-color: %s' % color1
+        elif val != 0.0 and is_last:
+            return 'background-color: %s' % color3
+    else:
+        return ''
+
+def calculate_percentage_difference(cur_array, previous_array):
+    new_array = []
+    for i in range(len(cur_array)):
+        if type(cur_array[i]) == type(pd.NA) or type(previous_array[i]) == type(pd.NA):
+            new_array.append(pd.NA)
+        else:
+            new_array.append(round((cur_array[i]-previous_array[i])*100/previous_array[i], 2))
+    return np.array(new_array)
+
+def check_diffs_within_normal_range(latest_csv, highlight_set, threshold):
+    within = True
+
+    for column in highlight_set:
+        for value in latest_csv[column]:
+            if type(value) != type(pd.NA):
+                within = within and abs(value) <= threshold
+    
+    return within
+
+
+def main():
+    parser = argparse.ArgumentParser(description="convert .csv file to .html file")
+    parser.add_argument("-f", "--folder_path", type=str, dest="folder_path",
+                        help="The directory which stores the .csv file", default="/home/arda/BigDL/python/llm/dev/benchmark/ceval")
+    parser.add_argument("-t", "--threshold", type=float, dest="threshold",
+                        help="the threshold of highlight values", default=3.0)
+    parser.add_argument("-b", "--baseline_path", type=str, dest="baseline_path",
+                        help="the baseline path which stores the baseline.csv file")
+    args = parser.parse_args()
+
+    csv_files = []
+    for file_name in os.listdir(args.folder_path):
+        file_path = os.path.join(args.folder_path, file_name)
+        if os.path.isfile(file_path) and file_name.endswith(".csv"):
+            csv_files.append(file_path)
+    csv_files.sort(reverse=True)
+
+    highlight_threshold=args.threshold
+    
+    # get the newest csv file
+    latest_csv = pd.read_csv(csv_files[0], index_col=0)
+
+    # create daily html file
+    daily_html=csv_files[0].split(".")[0]+".html"
+
+    # add index column
+    latest_csv.reset_index(inplace=True)
+
+    # if found more than 1 csv file
+    if len(csv_files)>1:
+        if args.baseline_path:
+            previous_csv = pd.read_csv(args.baseline_path, index_col=0)
+        else:
+            previous_csv = pd.read_csv(csv_files[1], index_col=0)
+
+        subjects = ['STEM', 'Social Science', 'Humanities', 'Other', 'Hard', 'Average']
+        precisions = ['sym_int4', 'fp8_e5m2']
+        highlight_set = []
+
+        insert_column = latest_csv.shape[-1]-1
+        # in the make_csv step we will handle the missing values and make it pd.NA
+        for subject in subjects:
+            # insert last accuracy task
+            latest_csv.insert(loc=insert_column, column=f'last_{subject}',
+                              value=previous_csv[subject])
+
+            # insert precentage difference between previous and current value
+            latest_csv.insert(
+                loc=insert_column+1,
+                column=f'diff_{subject}(%)',
+                value=calculate_percentage_difference(latest_csv[subject], previous_csv[subject]))
+            # append in the highlight set
+            highlight_set.append(f'diff_{subject}(%)')
+
+            # update insert column
+            insert_column += 2
+
+        columns = {}
+        for column in latest_csv.columns.values.tolist():
+            columns[column] = '{:.2f}'
+
+        styled_df = latest_csv.style.format(columns).applymap(lambda val: highlight_vals(val, max=3.0, is_last=True), subset=highlight_set)
+        
+        # add css style to restrict width and wrap text
+        styled_df.set_table_styles([{
+            'selector': 'th, td',
+            'props': [('max-width', '88px'), ('word-wrap', 'break-word')]
+        }], overwrite=False)
+        
+        html_output = styled_df.set_table_attributes("border=1").to_html()
+
+        with open(daily_html, 'w') as f:
+            f.write(html_output)
+    else:
+        latest_csv.to_html(daily_html)
+
+    if args.baseline_path and not check_diffs_within_normal_range(latest_csv, highlight_set, highlight_threshold):
+        print("The diffs are outside the normal range: %" + str(highlight_threshold))
+        return 1 
+    return 0
+
+if __name__ == "__main__":
+    sys.exit(main())

python/llm/test/benchmark/ceval/update_html_in_parent_folder.py

@@ -0,0 +1,49 @@
+#
+# 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.
+#
+
+# Python program to update Html in parent folder
+
+import os
+import shutil
+import argparse
+from pathlib import Path
+
+def update_html_in_parent_folder(folder_path):
+    # Get parent folder
+    parent_folder = Path(folder_path).parent
+
+    # List all html files under parent folder and delete them
+    for html_file in parent_folder.glob('*.html'):
+        html_file.unlink()
+
+    # Find latest html file under folder_path
+    latest_html_file = max(Path(folder_path).glob('*.html'), key=os.path.getctime, default=None)
+
+    # Copy the latest html file to parent folder
+    if latest_html_file is not None:
+        shutil.copy(latest_html_file, parent_folder)
+    
+    print(latest_html_file.name)
+
+def main():
+    parser = argparse.ArgumentParser(description="Update HTML in parent folder.")
+    parser.add_argument("-f", "--folder", type=str, help="Path to the folder")
+    args = parser.parse_args()
+
+    update_html_in_parent_folder(args.folder)
+
+if __name__ == "__main__":
+    main()