import json import os import pandas as pd from typing import List, Dict, Tuple from src.display.formatting import has_no_nan_values, make_clickable_model from src.display.utils import AutoEvalColumn, AutoEvalColumnMultimodal, EvalQueueColumn from src.leaderboard.read_evals import get_raw_eval_results, get_raw_eval_results_mib_subgraph, get_raw_eval_results_mib_causalgraph from src.about import TasksMib_Causalgraph def get_leaderboard_df(results_path: str, requests_path: str, cols: list, benchmark_cols: list) -> pd.DataFrame: """Creates a dataframe from all the individual experiment results""" # print(f"results_path is {results_path}, requests_path is {requests_path}") raw_data = get_raw_eval_results(results_path, requests_path) # print(f"raw_data is {raw_data}") all_data_json = [v.to_dict() for v in raw_data] # print(f"all_data_json is {pd.DataFrame.from_records(all_data_json)}") all_data_json_filtered = [] for item in all_data_json: item["Track"] = item["eval_name"].split("_")[-1] item["ioi"] = 0 item["mcqa"] = 0 if "VQA" in benchmark_cols and "VQA" in item: all_data_json_filtered.append(item) if "VQA" not in benchmark_cols and "VQA" not in item: all_data_json_filtered.append(item) all_data_json = all_data_json_filtered df = pd.DataFrame.from_records(all_data_json) df = df.sort_values(by=[AutoEvalColumn.text_average.name], ascending=False) # df = df.sort_values(by=[Tasks.task0.value.col_name], ascending=False) # df = df.sort_values(by=[AutoEvalColumn.track.name], ascending=False) # print(f"df is {df}") # df = df[cols].round(decimals=1) # filter out if any of the benchmarks have not been produced df = df[has_no_nan_values(df, benchmark_cols)] return df def get_leaderboard_df_mib_subgraph(results_path: str, requests_path: str, cols: list, benchmark_cols: list, metric_type = "F+") -> pd.DataFrame: """Creates a dataframe from all the MIB experiment results""" # print(f"results_path is {results_path}, requests_path is {requests_path}") raw_data = get_raw_eval_results_mib_subgraph(results_path, requests_path) # print(f"raw_data is {raw_data}") # Convert each result to dict format all_data_json = [v.to_dict(metric_type=metric_type) for v in raw_data] # print(f"all_data_json is {pd.DataFrame.from_records(all_data_json)}") # Convert to dataframe df = pd.DataFrame.from_records(all_data_json) # Sort by Average score descending if 'Average' in df.columns: # Convert '-' to NaN for sorting purposes df['Average'] = pd.to_numeric(df['Average'], errors='coerce') df = df.sort_values(by=['Average'], ascending=False, na_position='last') # Convert NaN back to '-' df['Average'] = df['Average'].fillna('-') return df # def aggregate_methods(df: pd.DataFrame) -> pd.DataFrame: # """Aggregates rows with the same base method name by taking the max value for each column""" # df_copy = df.copy() # # Extract base method names (remove _2, _3, etc. suffixes) # base_methods = [name.split('_')[0] if '_' in name and name.split('_')[-1].isdigit() # else name for name in df_copy.index] # df_copy.index = base_methods # # Convert scores to numeric values # numeric_df = df_copy.select_dtypes(include=['float64', 'int64']) # # Group by base method name and take the max # aggregated_df = numeric_df.groupby(level=0).max().round(3) # return aggregated_df def aggregate_methods(df: pd.DataFrame) -> pd.DataFrame: """Aggregates rows with the same base method name by taking the max value for each column""" df_copy = df.copy() # Set Method as index if it isn't already if 'Method' in df_copy.columns: df_copy.set_index('Method', inplace=True) # Extract base method names (remove _2, _3, etc. suffixes) base_methods = [name.split('_')[0] if '_' in str(name) and str(name).split('_')[-1].isdigit() else name for name in df_copy.index] df_copy.index = base_methods # Convert scores to numeric values numeric_df = df_copy.select_dtypes(include=['float64', 'int64']) # Group by base method name and take the max aggregated_df = numeric_df.groupby(level=0).max().round(3) # Reset index to get Method as a column aggregated_df.reset_index(inplace=True) aggregated_df.rename(columns={'index': 'Method'}, inplace=True) return aggregated_df # def create_intervention_averaged_df(df: pd.DataFrame) -> pd.DataFrame: # """Creates a DataFrame where columns are model_task and cells are averaged over interventions""" # df_copy = df.copy() # # Remove the Method column and eval_name if present # columns_to_drop = ['Method', 'eval_name'] # df_copy = df_copy.drop(columns=[col for col in columns_to_drop if col in df_copy.columns]) # # Group columns by model_task # model_task_groups = {} # for col in df_copy.columns: # model_task = '_'.join(col.split('_')[:2]) # Get model_task part # if model_task not in model_task_groups: # model_task_groups[model_task] = [] # model_task_groups[model_task].append(col) # # Create new DataFrame with averaged intervention scores # averaged_df = pd.DataFrame({ # model_task: df_copy[cols].mean(axis=1).round(3) # for model_task, cols in model_task_groups.items() # }) # return averaged_df # def create_intervention_averaged_df(df: pd.DataFrame) -> pd.DataFrame: # """Creates a DataFrame where columns are model_task and cells are averaged over interventions""" # df_copy = df.copy() # # Store Method column if it exists # method_col = None # if 'Method' in df_copy.columns: # method_col = df_copy['Method'] # df_copy = df_copy.drop('Method', axis=1) # # Remove eval_name if present # if 'eval_name' in df_copy.columns: # df_copy = df_copy.drop('eval_name', axis=1) # # Group columns by model_task # model_task_groups = {} # for col in df_copy.columns: # model_task = '_'.join(col.split('_')[:2]) # Get model_task part # if model_task not in model_task_groups: # model_task_groups[model_task] = [] # model_task_groups[model_task].append(col) # # Create new DataFrame with averaged intervention scores # averaged_df = pd.DataFrame({ # model_task: df_copy[cols].mean(axis=1).round(3) # for model_task, cols in model_task_groups.items() # }) # # Add Method column back # if method_col is not None: # averaged_df.insert(0, 'Method', method_col) # return averaged_df def create_intervention_averaged_df(df: pd.DataFrame) -> pd.DataFrame: """Creates a DataFrame where columns are model_task and cells are averaged over interventions""" df_copy = df.copy() # Store Method column method_col = None if 'Method' in df_copy.columns: method_col = df_copy['Method'] df_copy = df_copy.drop('Method', axis=1) if 'eval_name' in df_copy.columns: df_copy = df_copy.drop('eval_name', axis=1) # Group columns by model and intervention result_cols = {} for task in TasksMib_Causalgraph: for model in task.value.models: # Will iterate over all three models for intervention in task.value.interventions: for counterfactual in task.value.counterfactuals: col_pattern = f"{model}_layer.*_{intervention}_{counterfactual}" matching_cols = [c for c in df_copy.columns if pd.Series(c).str.match(col_pattern).any()] if matching_cols: col_name = f"{model}_{intervention}_{counterfactual}" result_cols[col_name] = matching_cols averaged_df = pd.DataFrame() if method_col is not None: averaged_df['Method'] = method_col for col_name, cols in result_cols.items(): averaged_df[col_name] = df_copy[cols].mean(axis=1).round(3) return averaged_df # def get_leaderboard_df_mib_causalgraph(results_path: str, requests_path: str, cols: list, benchmark_cols: list) -> pd.DataFrame: # """Creates a dataframe from all the MIB causal graph experiment results""" # print(f"results_path is {results_path}, requests_path is {requests_path}") # raw_data = get_raw_eval_results_mib_causalgraph(results_path, requests_path) # print(f"raw_data is {raw_data}") # # Convert each result to dict format for detailed df # all_data_json = [v.to_dict() for v in raw_data] # detailed_df = pd.DataFrame.from_records(all_data_json) # print(f"detailed_df is: {detailed_df}") # # Create and print other views for debugging/reference # aggregated_df = aggregate_methods(detailed_df) # print(f"aggregated_df is: {aggregated_df}") # intervention_averaged_df = create_intervention_averaged_df(aggregated_df) # print(f"intervention_averaged_df is: {intervention_averaged_df}") # # Only return detailed_df for display # return detailed_df # def get_leaderboard_df_mib_causalgraph(results_path: str, requests_path: str, cols: list, benchmark_cols: list) -> Tuple[pd.DataFrame, pd.DataFrame, pd.DataFrame]: # print(f"results_path is {results_path}, requests_path is {requests_path}") # raw_data = get_raw_eval_results_mib_causalgraph(results_path, requests_path) # # Convert each result to dict format for detailed df # all_data_json = [v.to_dict() for v in raw_data] # detailed_df = pd.DataFrame.from_records(all_data_json) # print("Columns in detailed_df:", detailed_df.columns.tolist()) # Print actual columns # # Create aggregated df # aggregated_df = aggregate_methods(detailed_df) # print("Columns in aggregated_df:", aggregated_df.columns.tolist()) # # Create intervention-averaged df # intervention_averaged_df = create_intervention_averaged_df(aggregated_df) # print("Columns in intervention_averaged_df:", intervention_averaged_df.columns.tolist()) # return detailed_df, aggregated_df, intervention_averaged_df def get_leaderboard_df_mib_causalgraph(results_path: str, requests_path: str, cols: list, benchmark_cols: list) -> Tuple[pd.DataFrame, pd.DataFrame, pd.DataFrame]: # print(f"results_path is {results_path}, requests_path is {requests_path}") raw_data = get_raw_eval_results_mib_causalgraph(results_path, requests_path) # Convert each result to dict format for detailed df all_data_json = [v.to_dict() for v in raw_data] detailed_df = pd.DataFrame.from_records(all_data_json) # Print the actual columns for debugging # print("Original columns:", detailed_df.columns.tolist()) # Rename columns to match schema column_mapping = {} for col in detailed_df.columns: if col in ['eval_name', 'Method']: continue # Ensure consistent casing for the column names new_col = col.replace('Qwen2ForCausalLM', 'qwen2forcausallm') \ .replace('Gemma2ForCausalLM', 'gemma2forcausallm') \ .replace('LlamaForCausalLM', 'llamaforcausallm') column_mapping[col] = new_col detailed_df = detailed_df.rename(columns=column_mapping) # Create aggregated df aggregated_df = aggregate_methods(detailed_df) # Create intervention-averaged df intervention_averaged_df = create_intervention_averaged_df(aggregated_df) # print("Transformed columns:", detailed_df.columns.tolist()) print(f"Columns in detailed_df: {detailed_df.columns.tolist()}") print(f"Columns in aggregated_df: {aggregated_df.columns.tolist()}") print(f"Columns in intervention_averaged_df: {intervention_averaged_df.columns.tolist()}") return detailed_df, aggregated_df, intervention_averaged_df def get_evaluation_queue_df(save_path: str, cols: list) -> list[pd.DataFrame]: """Creates the different dataframes for the evaluation queues requests""" entries = [entry for entry in os.listdir(save_path) if not entry.startswith(".")] all_evals = [] for entry in entries: if ".json" in entry: file_path = os.path.join(save_path, entry) with open(file_path) as fp: data = json.load(fp) if "still_on_hub" in data and data["still_on_hub"]: data[EvalQueueColumn.model.name] = make_clickable_model(data["hf_repo"], data["model"]) data[EvalQueueColumn.revision.name] = data.get("revision", "main") else: data[EvalQueueColumn.model.name] = data["model"] data[EvalQueueColumn.revision.name] = "N/A" all_evals.append(data) elif ".md" not in entry: # this is a folder sub_entries = [e for e in os.listdir(f"{save_path}/{entry}") if os.path.isfile(e) and not e.startswith(".")] for sub_entry in sub_entries: file_path = os.path.join(save_path, entry, sub_entry) with open(file_path) as fp: data = json.load(fp) data[EvalQueueColumn.model.name] = make_clickable_model(data["model"]) data[EvalQueueColumn.revision.name] = data.get("revision", "main") all_evals.append(data) pending_list = [e for e in all_evals if e["status"] in ["PENDING", "RERUN"]] running_list = [e for e in all_evals if e["status"] == "RUNNING"] finished_list = [e for e in all_evals if e["status"].startswith("FINISHED") or e["status"] == "PENDING_NEW_EVAL"] df_pending = pd.DataFrame.from_records(pending_list, columns=cols) df_running = pd.DataFrame.from_records(running_list, columns=cols) df_finished = pd.DataFrame.from_records(finished_list, columns=cols) return df_finished[cols], df_running[cols], df_pending[cols]