import gradio as gr import os from huggingface_hub import HfApi, snapshot_download from apscheduler.schedulers.background import BackgroundScheduler from datasets import load_dataset from src.utils import load_all_data from src.md import ABOUT_TEXT, TOP_TEXT from src.plt import plot_avg_correlation from src.constants import subset_mapping, length_categories, example_counts from src.css import custom_css import numpy as np api = HfApi() COLLAB_TOKEN = os.environ.get("COLLAB_TOKEN") evals_repo = "allenai/reward-bench-results" eval_set_repo = "allenai/reward-bench" repo_dir_rewardbench = "./evals/rewardbench/" def restart_space(): api.restart_space(repo_id="allenai/reward-bench", token=COLLAB_TOKEN) print("Pulling evaluation results") repo = snapshot_download( local_dir=repo_dir_rewardbench, ignore_patterns=["pref-sets-scores/*", "eval-set-scores/*"], repo_id=evals_repo, use_auth_token=COLLAB_TOKEN, tqdm_class=None, etag_timeout=30, repo_type="dataset", ) def avg_over_rewardbench(dataframe_core, dataframe_prefs): """ Averages over the subsets alpacaeval, mt-bench, llmbar, refusals, hep and returns dataframe with only these columns. We average over 4 core sections (per prompt weighting): 1. Chat: Includes the easy chat subsets (alpacaeval-easy, alpacaeval-length, alpacaeval-hard, mt-bench-easy, mt-bench-medium) 2. Chat Hard: Includes the hard chat subsets (mt-bench-hard, llmbar-natural, llmbar-adver-neighbor, llmbar-adver-GPTInst, llmbar-adver-GPTOut, llmbar-adver-manual) 3. Safety: Includes the safety subsets (refusals-dangerous, refusals-offensive, xstest-should-refuse, xstest-should-respond, do not answer) 4. Reasoning: Includes the code and math subsets (math-prm, hep-cpp, hep-go, hep-java, hep-js, hep-python, hep-rust) 5. Prior Sets (0.5 weight): Includes the test sets (anthropic_helpful, mtbench_human, shp, summarize) """ new_df = dataframe_core.copy() dataframe_prefs = dataframe_prefs.copy() # for main subsets, keys in subset_mapping, take the weighted avg by example_counts and store for the models for subset, sub_subsets in subset_mapping.items(): subset_cols = [col for col in new_df.columns if col in sub_subsets] sub_data = new_df[subset_cols].values # take the relevant column values sub_counts = [example_counts[s] for s in subset_cols] # take the example counts new_df[subset] = np.average(sub_data, axis=1, weights=sub_counts) # take the weighted average # new_df[subset] = np.round(np.nanmean(new_df[subset_cols].values, axis=1), 2) data_cols = list(subset_mapping.keys()) keep_columns = ["model",] + ["model_type"] + data_cols # keep_columns = ["model", "average"] + subsets new_df = new_df[keep_columns] # selected average from pref_sets pref_columns = ["anthropic_helpful", "anthropic_hhh", "shp", "summarize"] pref_data = dataframe_prefs[pref_columns].values # add column test sets knowing the rows are not identical, take superset dataframe_prefs["Prior Sets (0.5 weight)"] = np.nanmean(pref_data, axis=1) # add column Test Sets empty to new_df new_df["Prior Sets (0.5 weight)"] = np.nan # per row in new_df if model is in dataframe_prefs, add the value to new_df["Prior Sets (0.5 weight)"] values = [] for i, row in new_df.iterrows(): model = row["model"] if model in dataframe_prefs["model"].values: values.append(dataframe_prefs[dataframe_prefs["model"] == model]["Prior Sets (0.5 weight)"].values[0]) # new_df.at[i, "Prior Sets (0.5 weight)"] = dataframe_prefs[dataframe_prefs["model"] == model]["Prior Sets (0.5 weight)"].values[0] else: values.append(np.nan) new_df["Prior Sets (0.5 weight)"] = values # add total average data_cols += ["Prior Sets (0.5 weight)"] final_data = new_df[data_cols].values masked_data = np.ma.masked_array(final_data, np.isnan(final_data)) weights = [2, 2, 2, 2, 1] average = np.ma.average(masked_data, axis=1, weights=weights) new_df["average"] = average.filled(np.nan) # new_df["average"] = np.nanmean(new_df[data_cols].values, axis=1) # make average third column keep_columns = ["model", "model_type", "average"] + data_cols new_df = new_df[keep_columns] return new_df def expand_subsets(dataframe): # TODO need to modify data/ script to do this pass def length_bias_check(dataframe): """ Takes the raw rewardbench dataframe and splits the data into new buckets according to length_categories. Then, take the average of the three buckets as "average" """ new_df = dataframe.copy() existing_subsets = new_df.columns[3:] # model, model_type, average final_subsets = ["Length Bias", "Neutral", "Terse Bias"] # new data is empty list dict for each final subset new_data = {s: [] for s in final_subsets} # now, subsets correspond to those with True, Nuetral, and False length bias # check if length_categories[subset] == "True" or "False" or "Neutral" for subset in existing_subsets: subset_data = new_df[subset].values subset_length = length_categories[subset] # route to the correct bucket if subset_length == "True": new_data["Length Bias"].append(subset_data) elif subset_length == "Neutral": new_data["Neutral"].append(subset_data) elif subset_length == "False": new_data["Terse Bias"].append(subset_data) # take average of new_data and add to new_df (removing other columns than model) for subset in final_subsets: new_df[subset] = np.nanmean(new_data[subset], axis=0) keep_columns = ["model"] + final_subsets new_df = new_df[keep_columns] # recompute average # new_df["average"] = np.round(np.nanmean(new_df[final_subsets].values, axis=1), 2) return new_df rewardbench_data = load_all_data(repo_dir_rewardbench, subdir="eval-set").sort_values(by='average', ascending=False) rewardbench_data_length = length_bias_check(rewardbench_data).sort_values(by='Terse Bias', ascending=False) prefs_data = load_all_data(repo_dir_rewardbench, subdir="pref-sets").sort_values(by='average', ascending=False) # prefs_data_sub = expand_subsets(prefs_data).sort_values(by='average', ascending=False) rewardbench_data_avg = avg_over_rewardbench(rewardbench_data, prefs_data).sort_values(by='average', ascending=False) def prep_df(df): # add column to 0th entry with count (column name itself empty) df.insert(0, '', range(1, 1 + len(df))) # replace "model" with "Model" and "model_type" with "Model Type" and "average" with "Average" df = df.rename(columns={"model": "Model", "model_type": "Model Type", "average": "Average"}) # if "Model Type" in columns if "Model Type" in df.columns: # get model_types that have generative in them mask = df["Model Type"].str.contains("generative", case=False, na=False) # set these values to "Generative" df.loc[mask, "Model Type"] = "Generative" return df # add count column to all dataframes rewardbench_data = prep_df(rewardbench_data) rewardbench_data_avg = prep_df(rewardbench_data_avg).rename(columns={"Average": "Score"}) # adjust weight of this average to 50% for Prior Sets (0.5 weight), 1 for others rewardbench_data_length = prep_df(rewardbench_data_length) prefs_data = prep_df(prefs_data) col_types_rewardbench = ["number"] + ["markdown"] + ["str"] + ["number"] * (len(rewardbench_data.columns) - 1) col_types_rewardbench_avg = ["number"] + ["markdown"]+ ["str"] + ["number"] * (len(rewardbench_data_avg.columns) - 1) cols_rewardbench_data_length = ["markdown"] + ["number"] * (len(rewardbench_data_length.columns) - 1) col_types_prefs = ["number"] + ["markdown"] + ["number"] * (len(prefs_data.columns) - 1) # col_types_prefs_sub = ["markdown"] + ["number"] * (len(prefs_data_sub.columns) - 1) # for showing random samples eval_set = load_dataset(eval_set_repo, use_auth_token=COLLAB_TOKEN, split="filtered") def random_sample(r: gr.Request, subset): if subset is None or subset == []: sample_index = np.random.randint(0, len(eval_set) - 1) sample = eval_set[sample_index] else: # filter by subsets (can be list) if isinstance(subset, str): subset = [subset] # filter down dataset to only include the subset(s) eval_set_filtered = eval_set.filter(lambda x: x["subset"] in subset) sample_index = np.random.randint(0, len(eval_set_filtered) - 1) sample = eval_set_filtered[sample_index] markdown_text = '\n\n'.join([f"**{key}**:\n\n{value}" for key, value in sample.items()]) return markdown_text subsets = eval_set.unique("subset") def regex_table(dataframe, regex, filter_button): """ Takes a model name as a regex, then returns only the rows that has that in it. """ # Split regex statement by comma and trim whitespace around regexes regex_list = [x.strip() for x in regex.split(",")] # Join the list into a single regex pattern with '|' acting as OR combined_regex = '|'.join(regex_list) # remove internal ai2 data dataframe = dataframe[~dataframe["Model"].str.contains("ai2", case=False, na=False)] # if filter_button, remove all rows with "ai2" in the model name update_scores = False if isinstance(filter_button, list) or isinstance(filter_button, str): if "Prior Sets" not in filter_button and 'Prior Sets (0.5 weight)' in dataframe.columns: update_scores = True if "Seq. Classifiers" not in filter_button: dataframe = dataframe[~dataframe["Model Type"].str.contains("Seq. Classifier", case=False, na=False)] if "DPO" not in filter_button: dataframe = dataframe[~dataframe["Model Type"].str.contains("DPO", case=False, na=False)] if "Custom Classifiers" not in filter_button: dataframe = dataframe[~dataframe["Model Type"].str.contains("Custom Classifier", case=False, na=False)] if "Generative" not in filter_button: dataframe = dataframe[~dataframe["Model Type"].str.contains("generative", case=False, na=False)] # Filter the dataframe such that 'model' contains any of the regex patterns data = dataframe[dataframe["Model"].str.contains(combined_regex, case=False, na=False)] # replace column '' with count/rank data[''] = np.arange(1, 1 + len(data)) # if update the score to not use prior sets, do so if update_scores: data["Score"] = (data["Chat"] + data["Chat Hard"] + data["Safety"] + data["Reasoning"]) / 4 data["Prior Sets (0.5 weight)"] = np.NaN # sort array by Score column data = data.sort_values(by='Score', ascending=False) # if Score exists, round to 2 decimals if "Score" in data.columns: data["Score"] = np.round(np.array(data["Score"].values).astype(float), 2) if "Average" in data.columns: data["Average"] = np.round(np.array(data["Average"].values).astype(float), 1) # round all others to 1 decimal for col in data.columns: if col not in ["", "Model", "Model Type", "Score", "Average"]: # replace any data[col].values == '' with np.NaN data[col] = data[col].replace('', np.NaN) data[col] = np.round(np.array(data[col].values).astype(float), 1) return data # import ipdb; ipdb.set_trace() total_models = len(regex_table(rewardbench_data_avg.copy(), "", ["Seq. Classifiers", "DPO", "Custom Classifiers", "Generative"]).values) with gr.Blocks(css=custom_css) as app: # create tabs for the app, moving the current table to one titled "rewardbench" and the benchmark_text to a tab called "About" with gr.Row(): with gr.Column(scale=6): gr.Markdown(TOP_TEXT.format(str(total_models))) with gr.Column(scale=4): # search = gr.Textbox(label="Model Search (delimit with , )", placeholder="Regex search for a model") # filter_button = gr.Checkbox(label="Include AI2 training runs (or type ai2 above).", interactive=True) # img = gr.Image(value="https://private-user-images.githubusercontent.com/10695622/310698241-24ed272a-0844-451f-b414-fde57478703e.png", width=500) gr.Markdown(""" ![](file/src/logo.png) """) with gr.Tabs(elem_classes="tab-buttons") as tabs: with gr.TabItem("🏆 RewardBench Leaderboard"): with gr.Row(): search_1 = gr.Textbox(label="Model Search (delimit with , )", placeholder="Model Search (delimit with , )", show_label=False) model_types_1 = gr.CheckboxGroup(["Seq. Classifiers", "DPO", "Custom Classifiers", "Generative", "Prior Sets"], value=["Seq. Classifiers", "DPO", "Custom Classifiers", "Prior Sets"], label="Model Types", show_label=False, # info="Which model types to include.", ) with gr.Row(): # reference data rewardbench_table_hidden = gr.Dataframe( rewardbench_data_avg.values, datatype=col_types_rewardbench_avg, headers=rewardbench_data_avg.columns.tolist(), visible=False, ) rewardbench_table = gr.Dataframe( regex_table(rewardbench_data_avg.copy(), "", ["Seq. Classifiers", "DPO", "Custom Classifiers", "Prior Sets"]).values, datatype=col_types_rewardbench_avg, headers=rewardbench_data_avg.columns.tolist(), elem_id="rewardbench_dataframe_avg", height=1000, ) with gr.TabItem("🔍 RewardBench - Detailed"): with gr.Row(): search_2 = gr.Textbox(label="Model Search (delimit with , )", show_label=False, placeholder="Model Search (delimit with , )") model_types_2 = gr.CheckboxGroup(["Seq. Classifiers", "DPO", "Custom Classifiers", "Generative"], value=["Seq. Classifiers", "DPO", "Custom Classifiers"], label="Model Types", show_label=False, # info="Which model types to include." ) with gr.Row(): # ref data rewardbench_table_detailed_hidden = gr.Dataframe( rewardbench_data.values, datatype=col_types_rewardbench, headers=rewardbench_data.columns.tolist(), visible=False, ) rewardbench_table_detailed = gr.Dataframe( regex_table(rewardbench_data.copy(), "", ["Seq. Classifiers", "DPO", "Custom Classifiers"]).values, datatype=col_types_rewardbench, headers=rewardbench_data.columns.tolist(), elem_id="rewardbench_dataframe", height=1000, ) # with gr.TabItem("rewardbench Eval Set - Length Bias"): # with gr.Row(): # # backup # rewardbench_table_len_hidden = gr.Dataframe( # rewardbench_data_length.values, # datatype=cols_rewardbench_data_length, # headers=rewardbench_data_length.columns.tolist(), # visible=False, # ) # rewardbench_table_len = gr.Dataframe( # regex_table(rewardbench_data_length.copy(), "", False).values, # datatype=cols_rewardbench_data_length, # headers=rewardbench_data_length.columns.tolist(), # elem_id="rewardbench_dataframe_length", # height=1000, # ) with gr.TabItem("Prior Test Sets"): with gr.Row(): search_3 = gr.Textbox(label="Model Search (delimit with , )", show_label=False, placeholder="Model Search (delimit with , )") model_types_3 = gr.CheckboxGroup(["Seq. Classifiers", "DPO", "Custom Classifiers", "Generative"], value=["Seq. Classifiers", "DPO", "Custom Classifiers"], label="Model Types", show_label=False, # info="Which model types to include.", ) with gr.Row(): PREF_SET_TEXT = """ For more information, see the [dataset](https://huggingface.co/datasets/allenai/pref-test-sets). Only the subsets Anthropic Helpful, Anthropic HHH, Stanford SHP, and OpenAI's Summarize data are used in the leaderboard ranking. """ gr.Markdown(PREF_SET_TEXT) with gr.Row(): # backup pref_sets_table_hidden = gr.Dataframe( prefs_data.values, datatype=col_types_prefs, headers=prefs_data.columns.tolist(), visible=False, ) pref_sets_table = gr.Dataframe( regex_table(prefs_data.copy(), "", ["Seq. Classifiers", "DPO", "Custom Classifiers"]).values, datatype=col_types_prefs, headers=prefs_data.columns.tolist(), elem_id="prefs_dataframe", height=1000, ) with gr.TabItem("About"): with gr.Row(): gr.Markdown(ABOUT_TEXT) with gr.TabItem("Dataset Viewer"): with gr.Row(): # loads one sample gr.Markdown("""## Random Dataset Sample Viewer Warning, refusals, XSTest, and donotanswer datasets have sensitive content.""") subset_selector = gr.Dropdown(subsets, label="Subset", value=None, multiselect=True) button = gr.Button("Show Random Sample") with gr.Row(): sample_display = gr.Markdown("{sampled data loads here}") button.click(fn=random_sample, inputs=[subset_selector], outputs=[sample_display]) # removed plot because not pretty enough # with gr.TabItem("Model Correlation"): # with gr.Row(): # plot = plot_avg_correlation(rewardbench_data_avg, prefs_data) # gr.Plot(plot) search_1.change(regex_table, inputs=[rewardbench_table_hidden, search_1, model_types_1], outputs=rewardbench_table) search_2.change(regex_table, inputs=[rewardbench_table_detailed_hidden, search_2, model_types_2], outputs=rewardbench_table_detailed) # search.change(regex_table, inputs=[rewardbench_table_len_hidden, search, filter_button], outputs=rewardbench_table_len) search_3.change(regex_table, inputs=[pref_sets_table_hidden, search_3, model_types_3], outputs=pref_sets_table) model_types_1.change(regex_table, inputs=[rewardbench_table_hidden, search_1, model_types_1], outputs=rewardbench_table) model_types_2.change(regex_table, inputs=[rewardbench_table_detailed_hidden, search_2, model_types_2], outputs=rewardbench_table_detailed) model_types_3.change(regex_table, inputs=[pref_sets_table_hidden, search_3, model_types_3], outputs=pref_sets_table) with gr.Row(): with gr.Accordion("📚 Citation", open=False): citation_button = gr.Textbox( value=r"""@misc{RewardBench, title={RewardBench: Evaluating Reward Models for Language Modeling}, author={Lambert, Nathan and Pyatkin, Valentina and Morrison, Jacob and Miranda, LJ and Lin, Bill Yuchen and Chandu, Khyathi and Dziri, Nouha and Kumar, Sachin and Zick, Tom and Choi, Yejin and Smith, Noah A. and Hajishirzi, Hannaneh}, year={2024}, howpublished={\url{https://huggingface.co/spaces/allenai/reward-bench} }""", lines=7, label="Copy the following to cite these results.", elem_id="citation-button", show_copy_button=True, ) # Load data when app starts, TODO make this used somewhere... # def load_data_on_start(): # data_rewardbench = load_all_data(repo_dir_rewardbench) # rewardbench_table.update(data_rewardbench) # data_rewardbench_avg = avg_over_rewardbench(repo_dir_rewardbench) # rewardbench_table.update(data_rewardbench_avg) # data_prefs = load_all_data(repo_dir_prefs) # pref_sets_table.update(data_prefs) scheduler = BackgroundScheduler() scheduler.add_job(restart_space, "interval", seconds=10800) # restarted every 3h scheduler.start() app.launch(allowed_paths=['src/']) # had .queue() before launch before... not sure if that's necessary