Merge pull request #10697 from MargarettMao/ceval

combine english and chinese, remove nan

python/llm/dev/benchmark/perplexity/make_csv.py

@@ -35,9 +35,8 @@ def make_csv(result_dict, output_path=None):
     current_date = datetime.datetime.now().strftime("%Y-%m-%d")
     file_name = f'results_{current_date}.csv'
     full_path = os.path.join(output_path, file_name) if output_path else file_name
-    print('Writing to', full_path)
     file_name = full_path
-    headers = ["Index", "Model", "Precision", "en", "zh"]
+    headers = ["Index", "Model", "Precision", "ppl_result"]
 
     with open(file_name, mode='w', newline='') as csv_file:
         writer = csv.writer(csv_file)
@@ -46,10 +45,10 @@ def make_csv(result_dict, output_path=None):
         for model, model_results in result_dict.items():
             for precision, prec_results in model_results.items():
                 row = [index, model, precision]
-                for language in headers[3:]:
+                for language in ["en","zh"]:
                     task_results = prec_results.get(language.lower(), None)
                     if task_results is None:
-                        row.append("")
+                        continue
                     else:
                         result = task_results["results"]
                         row.append("%.4f" % result)
@@ -89,6 +88,7 @@ def main(*args):
     output_path = args[2]
 
     merged_results = merge_results(input_path)
+
     make_csv(merged_results, output_path)
 
 

python/llm/dev/benchmark/perplexity/make_table.py

@@ -35,8 +35,8 @@ def make_table(result_dict):
     """Generate table of results."""
     md_writer = MarkdownTableWriter()
     latex_writer = LatexTableWriter()
-    md_writer.headers = ["Model", "Precision", "en", "zh"]
-    latex_writer.headers = ["Model", "Precision", "en", "zh"]
+    md_writer.headers =  ["Model", "Precision", "ppl_result"]
+    latex_writer.headers = ["Model", "Precision", "ppl_result"]
 
     languages = ["en", "zh"]
     values = []
@@ -46,7 +46,7 @@ def make_table(result_dict):
             for language in languages:
                 task_results = prec_results.get(language, None)
                 if task_results is None:
-                    value.append("")
+                    continue
                 else:
                     result = task_results["results"]
                     value.append("%.4f" % result)

python/llm/test/benchmark/perplexity/fp16.csv

@@ -1,8 +1,8 @@
-Index,Model,Precision,en,zh
-0,Llama-2-7b-chat-hf,fp16,4.7019,
-1,chatglm2-6b,fp16,,22.321
-2,chatglm3-6b,fp16,,30.1281
-3,Baichuan2-7B-Chat,fp16,,10.7676
-4,mpt-7b-chat,fp16,5.7882,
-5,falcon-7b-instruct-with-patch,fp16,5.2532,
-6,mistral-7b-v0.1,fp16,3.6597,
+Index,Model,Precision,ppl_result
+0,Llama-2-7b-chat-hf,fp16,4.7019
+1,chatglm2-6b,fp16,22.321
+2,chatglm3-6b,fp16,30.1281
+3,Baichuan2-7B-Chat,fp16,10.7676
+4,mpt-7b-chat,fp16,5.7882
+5,falcon-7b-instruct-with-patch,fp16,5.2532
+6,Mistral-7B-v0.1,fp16,3.6597

python/llm/test/benchmark/perplexity/ppl_csv_to_html.py

@@ -28,7 +28,7 @@ def highlight_vals(val, max=3.0, color1='red', color2='green', color3='yellow',
             return 'background-color: %s' % color1
         elif val <= -max:
             return 'background-color: %s' % color2
-        elif val != 0.0 and not pd.isna(val) and is_last:
+        elif val != 0.0 and is_last:
             return 'background-color: %s' % color3
     else:
         return ''
@@ -37,8 +37,8 @@ def nonzero_min(lst):
     non_zero_lst = [num for num in lst if num > 0.0]
     return min(non_zero_lst) if non_zero_lst else None
 
-def is_diffs_within_normal_range(diff_en, diff_zh, threshold=5.0):
-    return not any(diff < (-threshold) for diff in diff_en + diff_zh if isinstance(diff, float))
+def is_diffs_within_normal_range(diff_ppl_result, threshold=5.0):
+    return not any(diff < (-threshold) for diff in diff_ppl_result if isinstance(diff, float))
 
 def add_to_dict(dict, key, value):
     if key not in dict:
@@ -60,9 +60,9 @@ def create_fp16_dict(fp16_path):
         model = row['Model']
         # Formalize the data to have 2 decimal places
         fp16_dict[model] = {
-            'en': "{:.2f}".format(row['en']),
-            'zh': "{:.2f}".format(row['zh'])
+            'ppl_result': "{:.2f}".format(row['ppl_result'])
         }
+
     return fp16_dict
 
 def calculate_percentage_difference(current, fp16):
@@ -105,41 +105,34 @@ def main():
     diffs_within_normal_range = True
 
     # Add display of FP16 values for each model and add percentage difference column
-    for task in ['en', 'zh']:
-        latest_csv[f'{task}_FP16'] = latest_csv['Model'].apply(lambda model: fp16_dict.get(model, {}).get(task, 'N/A'))
-        latest_csv[f'{task}_diff_FP16(%)'] = latest_csv.apply(lambda row: calculate_percentage_difference(row[task], row[f'{task}_FP16']), axis=1)
+    latest_csv['ppl_result_FP16'] = latest_csv['Model'].apply(lambda model: fp16_dict.get(model, {}).get('ppl_result', 'N/A'))
+    latest_csv['ppl_result_diff_FP16(%)'] = latest_csv.apply(lambda row: calculate_percentage_difference(row['ppl_result'], row['ppl_result_FP16']), axis=1)
 
     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)
 
-        last_en=['']*len(latest_csv.index)
-        diff_en=['']*len(latest_csv.index)
-        last_zh=['']*len(latest_csv.index)
-        diff_zh=['']*len(latest_csv.index)
+        last_ppl_result=['']*len(latest_csv.index)
+        diff_ppl_result=['']*len(latest_csv.index)
 
-        en='en'
-        zh='zh'
+        ppl_result = 'ppl_result'
 
         csv_dict = {}
         for csv_file in csv_files:
             current_csv = pd.read_csv(csv_file, index_col=0)
             for current_csv_ind,current_csv_row in current_csv.iterrows():
                 current_csv_model=current_csv_row['Model'].strip()
                 current_csv_precision=current_csv_row['Precision'].strip()
-                current_csv_model_en=current_csv_model+'-'+current_csv_precision+'-'+'en'
-                current_csv_model_zh=current_csv_model+'-'+current_csv_precision+'-'+'zh'
-                add_to_dict(csv_dict, current_csv_model_en, current_csv_row[en])
-                add_to_dict(csv_dict, current_csv_model_zh, current_csv_row[zh]) 
-
+                current_csv_model_ppl_result=current_csv_model+'-'+current_csv_precision+'-'+'ppl_result'
+                add_to_dict(csv_dict, current_csv_model_ppl_result, current_csv_row[ppl_result])
+
         for latest_csv_ind,latest_csv_row in latest_csv.iterrows():
 
             latest_csv_model=latest_csv_row['Model'].strip()
             latest_csv_precision=latest_csv_row['Precision'].strip()
-            latest_en=latest_csv_row[en]
-            latest_zh=latest_csv_row[zh]
+            latest_ppl_result=latest_csv_row[ppl_result]
 
             in_previous_flag=False
 
@@ -150,57 +143,50 @@ def main():
 
                 if latest_csv_model==previous_csv_model and latest_csv_precision==previous_csv_precision:
 
-                    previous_en=previous_csv_row[en]
-                    previous_zh=previous_csv_row[zh]
-                    if previous_en > 0.0 or previous_zh > 0.0:
-                        last_en[latest_csv_ind]=previous_en
-                        diff_en[latest_csv_ind]=round((latest_en-previous_en)*100/previous_en,2)
-                        last_zh[latest_csv_ind]=previous_zh
-                        diff_zh[latest_csv_ind]=round((latest_zh-previous_zh)*100/previous_zh,2)
+                    previous_ppl_result=previous_csv_row[ppl_result] 
+
+                    if previous_ppl_result > 0.0:
+                        last_ppl_result[latest_csv_ind]=previous_ppl_result
+                        diff_ppl_result[latest_csv_ind]=round((latest_ppl_result-previous_ppl_result)*100/previous_ppl_result,2)
                         in_previous_flag=True
 
             if not in_previous_flag:
-                last_en[latest_csv_ind]=pd.NA
-                diff_en[latest_csv_ind]=pd.NA
-                last_zh[latest_csv_ind]=pd.NA
-                diff_zh[latest_csv_ind]=pd.NA
+                last_ppl_result[latest_csv_ind]=pd.NA
+                diff_ppl_result[latest_csv_ind]=pd.NA
 
-        latest_csv.insert(loc=9,column='last_en',value=last_en)
-        latest_csv.insert(loc=10,column='diff_en(%)',value=diff_en)
-        latest_csv.insert(loc=11,column='last_zh',value=last_zh)
-        latest_csv.insert(loc=12,column='diff_zh(%)',value=diff_zh)
 
+        latest_csv.insert(loc=6,column='last_ppl_result',value=last_ppl_result)
+        latest_csv.insert(loc=7,column='ppl_result_diff_last(%)',value=diff_ppl_result)
 
-        diffs_within_normal_range = is_diffs_within_normal_range(diff_en, diff_zh, threshold=highlight_threshold)
 
-        subset1=['diff_en(%)','diff_zh(%)']
+        diffs_within_normal_range = is_diffs_within_normal_range(diff_ppl_result, threshold=highlight_threshold)
 
-        columns={'en': '{:.2f}', 'zh': '{:.2f}', 'last_en': '{:.2f}', 'diff_en(%)': '{:.2f}',
-                'last_zh': '{:.2f}', 'diff_zh(%)': '{:.2f}'}
-
+        columns={'ppl_result': '{:.2f}', 'last_ppl_result': '{:.2f}', 'ppl_result_diff_last(%)': '{:.2f}'}
         latest_csv.drop('Index', axis=1, inplace=True)
 
-        styled_df = latest_csv.style.format(columns).applymap(lambda val: highlight_vals(val, max=3.0, is_last=True), subset=subset1)
-        for task in ['en', 'zh']:
-            styled_df = styled_df.applymap(lambda val: highlight_vals(val, max=highlight_threshold, is_last=False), subset=[f'{task}_diff_FP16(%)'])
-
-        # 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)
+        styled_df = latest_csv.style.format(columns).applymap(lambda val: highlight_vals(val, max=highlight_threshold, is_last=True), subset=['ppl_result_diff_last(%)'])
+        styled_df = styled_df.applymap(lambda val: highlight_vals(val, max=highlight_threshold, is_last=False), subset=['ppl_result_diff_FP16(%)'])
+
     else:
-        latest_csv.to_html(daily_html)
-
+        columns={'ppl_result': '{:.2f}'}
+        latest_csv.drop('Index', axis=1, inplace=True)
+        styled_df = latest_csv.style.format(columns).applymap(lambda val: highlight_vals(val, max=highlight_threshold, is_last=False), subset=['ppl_result_diff_FP16(%)'])
+
+    # 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)
+
     if args.baseline_path and not diffs_within_normal_range:
         print("The diffs are outside the normal range: %" + str(highlight_threshold))
-        return 1 
+        return 1
     return 0
-
+ 
 if __name__ == "__main__":
     sys.exit(main())