import subprocess import gradio as gr import pandas as pd from apscheduler.schedulers.background import BackgroundScheduler from huggingface_hub import snapshot_download from src.about import ( CITATION_BUTTON_LABEL, CITATION_BUTTON_TEXT, EVALUATION_QUEUE_TEXT, INTRODUCTION_TEXT, LLM_BENCHMARKS_TEXT, TITLE, ) from src.display.css_html_js import custom_css from src.display.utils import ( BENCHMARK_COLS, COLS, EVAL_COLS, EVAL_TYPES, NUMERIC_INTERVALS, TYPES, AutoEvalColumn, ModelType, fields, WeightType, Precision ) from src.envs import API, EVAL_REQUESTS_PATH, EVAL_RESULTS_PATH, QUEUE_REPO, REPO_ID, RESULTS_REPO, TOKEN from src.populate import get_evaluation_queue_df, get_leaderboard_df from src.submission.submit import add_new_eval def restart_space(): API.restart_space(repo_id=REPO_ID) try: print(EVAL_REQUESTS_PATH) snapshot_download( repo_id=QUEUE_REPO, local_dir=EVAL_REQUESTS_PATH, repo_type="dataset", tqdm_class=None, etag_timeout=30, token=TOKEN ) except Exception: restart_space() try: print(EVAL_RESULTS_PATH) snapshot_download( repo_id=RESULTS_REPO, local_dir=EVAL_RESULTS_PATH, repo_type="dataset", tqdm_class=None, etag_timeout=30, token=TOKEN ) except Exception: restart_space() raw_data, original_df = get_leaderboard_df(EVAL_RESULTS_PATH, EVAL_REQUESTS_PATH, COLS, BENCHMARK_COLS) leaderboard_df = original_df.copy() ( finished_eval_queue_df, running_eval_queue_df, pending_eval_queue_df, ) = get_evaluation_queue_df(EVAL_REQUESTS_PATH, EVAL_COLS) # Searching and filtering def update_table( hidden_df: pd.DataFrame, columns: list, type_query: list, precision_query: str, size_query: list, show_deleted: bool, italian_only: bool, query: str, ): filtered_df = filter_models(hidden_df, type_query, size_query, precision_query, show_deleted, italian_only) filtered_df = filter_queries(query, filtered_df) df = select_columns(filtered_df, columns) return df def search_table(df: pd.DataFrame, query: str) -> pd.DataFrame: return df[(df[AutoEvalColumn.model.name].str.contains(query, case=False))] def select_columns(df: pd.DataFrame, columns: list) -> pd.DataFrame: always_here_cols = [ AutoEvalColumn.model_type_symbol.name, AutoEvalColumn.model.name, ] # We use COLS to maintain sorting filtered_df = df[ always_here_cols + [c for c in COLS if c in df.columns and c in columns] ] return filtered_df def filter_queries(query: str, filtered_df: pd.DataFrame) -> pd.DataFrame: final_df = [] if query != "": queries = [q.strip() for q in query.split(";")] for _q in queries: _q = _q.strip() if _q != "": temp_filtered_df = search_table(filtered_df, _q) if len(temp_filtered_df) > 0: final_df.append(temp_filtered_df) if len(final_df) > 0: filtered_df = pd.concat(final_df) filtered_df = filtered_df.drop_duplicates( subset=[AutoEvalColumn.model.name, AutoEvalColumn.precision.name, AutoEvalColumn.revision.name] ) return filtered_df # def filter_models( # df: pd.DataFrame, type_query: list, size_query: list, precision_query: list, show_deleted: bool # ) -> pd.DataFrame: # # Show all models # if show_deleted: # filtered_df = df # else: # Show only still on the hub models # filtered_df = df[df[AutoEvalColumn.still_on_hub.name] == True] # type_emoji = [t[0] for t in type_query] # filtered_df = filtered_df.loc[df[AutoEvalColumn.model_type_symbol.name].isin(type_emoji)] # filtered_df = filtered_df.loc[df[AutoEvalColumn.precision.name].isin(precision_query + ["None"])] # numeric_interval = pd.IntervalIndex(sorted([NUMERIC_INTERVALS[s] for s in size_query])) # params_column = pd.to_numeric(df[AutoEvalColumn.params.name], errors="coerce") # mask = params_column.apply(lambda x: any(numeric_interval.contains(x))) # filtered_df = filtered_df.loc[mask] # return filtered_df def filter_models( df: pd.DataFrame, type_query: list, size_query: list, precision_query: list, show_deleted: bool, italian_only: bool ) -> pd.DataFrame: # Show all models if show_deleted: filtered_df = df.copy() else: # Show only still on the hub models filtered_df = df[df[AutoEvalColumn.still_on_hub.name] == True].copy() type_emoji = [t[0] for t in type_query] filtered_df = filtered_df[filtered_df[AutoEvalColumn.model_type_symbol.name].isin(type_emoji)] filtered_df = filtered_df[filtered_df[AutoEvalColumn.precision.name].isin(precision_query + ["None"])] numeric_interval = pd.IntervalIndex(sorted([NUMERIC_INTERVALS[s] for s in size_query])) params_column = pd.to_numeric(filtered_df[AutoEvalColumn.params.name], errors="coerce") mask = params_column.apply(lambda x: any(numeric_interval.contains(x))) filtered_df = filtered_df[mask] if italian_only: filtered_df = filtered_df[filtered_df[AutoEvalColumn.author.name] == "๐Ÿ‡ฎ๐Ÿ‡น"] return filtered_df def get_data_totale(): dataset = pd.read_csv("mmlu_pro_it.csv", sep=',') if 'model ' in dataset.columns: dataset.rename(columns={'model ': 'model'}, inplace=True) return dataset demo = gr.Blocks(css=custom_css) with demo: gr.HTML(TITLE) gr.Markdown(INTRODUCTION_TEXT, elem_classes="markdown-text") with gr.Tabs(elem_classes="tab-buttons") as tabs: with gr.TabItem("๐Ÿ… LLM Benchmark", elem_id="llm-benchmark-tab-table", id=0): with gr.Row(): with gr.Column(): with gr.Row(): search_bar = gr.Textbox( placeholder=" ๐Ÿ” Search for your model (separate multiple queries with `;`) and press ENTER...", show_label=False, elem_id="search-bar", ) with gr.Row(): shown_columns = gr.CheckboxGroup( choices=[ c.name for c in fields(AutoEvalColumn) if not c.hidden and not c.never_hidden ], value=[ c.name for c in fields(AutoEvalColumn) if c.displayed_by_default and not c.hidden and not c.never_hidden ], label="Select columns to show", elem_id="column-select", interactive=True, ) with gr.Row(): deleted_models_visibility = gr.Checkbox( value=True, label="Show gated/private/deleted models", interactive=True ) with gr.Column(min_width=320): italian_only_checkbox = gr.Checkbox( label="Show only models made by Italians", value=False, interactive=True, ) filter_columns_type = gr.CheckboxGroup( label="Model types", choices=[t.to_str() for t in ModelType], value=[t.to_str() for t in ModelType], interactive=True, elem_id="filter-columns-type", ) filter_columns_precision = gr.CheckboxGroup( label="Precision", choices=[i.value.name for i in Precision], value=[i.value.name for i in Precision], interactive=True, elem_id="filter-columns-precision", ) filter_columns_size = gr.CheckboxGroup( label="Model sizes (in billions of parameters)", choices=list(NUMERIC_INTERVALS.keys()), value=list(NUMERIC_INTERVALS.keys()), interactive=True, elem_id="filter-columns-size", ) leaderboard_table = gr.components.Dataframe( value=leaderboard_df[ [c.name for c in fields(AutoEvalColumn) if c.never_hidden] + shown_columns.value ], headers=[c.name for c in fields(AutoEvalColumn) if c.never_hidden] + shown_columns.value, datatype=TYPES, elem_id="leaderboard-table", interactive=False, visible=True, ) # Dummy leaderboard for handling the case when the user uses backspace key hidden_leaderboard_table_for_search = gr.components.Dataframe( value=original_df[COLS], headers=COLS, datatype=TYPES, visible=False, ) search_bar.submit( update_table, [ hidden_leaderboard_table_for_search, shown_columns, filter_columns_type, filter_columns_precision, filter_columns_size, deleted_models_visibility, italian_only_checkbox, search_bar, ], leaderboard_table, ) for selector in [shown_columns, filter_columns_type, filter_columns_precision, filter_columns_size, deleted_models_visibility, italian_only_checkbox]: selector.change( update_table, [ hidden_leaderboard_table_for_search, shown_columns, filter_columns_type, filter_columns_precision, filter_columns_size, deleted_models_visibility, italian_only_checkbox, search_bar, ], leaderboard_table, queue=True, ) # with gr.TabItem('Classifica RAG'): # gr.Markdown('''# Classifica RAG degli LLM italiani''') # gr.Markdown(f'''In questa sezione i modelli sono valutati su dei task di Q&A e ordinati per F1 Score e EM (Exact Match). La repo di riferimento รจ [questa](https://github.com/C080/open-llm-ita-leaderboard). # I modelli in cima alla classifica sono ritenuti preferibili per i task di Retrieval Augmented Generation.''') # gr.Dataframe(pd.read_csv('leaderboard.csv', sep=';')) # gr.Markdown(f"Si ringrazia il @galatolo per il codice dell'eval.") with gr.TabItem('Eval aggiuntive'): gr.Markdown('''# Altre evaluation''') gr.Markdown('''* classifica [INVALSI](https://huggingface.co/spaces/Crisp-Unimib/INVALSIbenchmark) gestita dai nostri amici del [CRISP](https://crispresearch.it/)''') gr.Markdown('''* analisi dei modelli fatti da ita su [mmlu pro it](https://huggingface.co/datasets/efederici/MMLU-Pro-ita)''') gr.Dataframe(get_data_totale) with gr.TabItem("๐Ÿ“ About", elem_id="llm-benchmark-tab-table", id=2): gr.Markdown(LLM_BENCHMARKS_TEXT, elem_classes="markdown-text") with gr.TabItem("๐Ÿš€ Submit here! ", elem_id="llm-benchmark-tab-table", id=3): with gr.Column(): with gr.Row(): gr.Markdown(EVALUATION_QUEUE_TEXT, elem_classes="markdown-text") with gr.Column(): with gr.Accordion( f"โœ… Finished Evaluations ({len(finished_eval_queue_df)})", open=False, ): with gr.Row(): finished_eval_table = gr.components.Dataframe( value=finished_eval_queue_df, headers=EVAL_COLS, datatype=EVAL_TYPES, row_count=5, ) with gr.Accordion( f"๐Ÿ”„ Running Evaluation Queue ({len(running_eval_queue_df)})", open=False, ): with gr.Row(): running_eval_table = gr.components.Dataframe( value=running_eval_queue_df, headers=EVAL_COLS, datatype=EVAL_TYPES, row_count=5, ) with gr.Accordion( f"โณ Pending Evaluation Queue ({len(pending_eval_queue_df)})", open=False, ): with gr.Row(): pending_eval_table = gr.components.Dataframe( value=pending_eval_queue_df, headers=EVAL_COLS, datatype=EVAL_TYPES, row_count=5, ) with gr.Row(): gr.Markdown("# โœ‰๏ธโœจ Submit your model here!", elem_classes="markdown-text") with gr.Row(): with gr.Column(): model_name_textbox = gr.Textbox(label="Model name") revision_name_textbox = gr.Textbox(label="Revision commit", placeholder="main") model_type = gr.Dropdown( choices=[t.to_str(" : ") for t in ModelType if t != ModelType.Unknown], label="Model type", multiselect=False, value=None, interactive=True, ) with gr.Column(): precision = gr.Dropdown( choices=[i.value.name for i in Precision if i != Precision.Unknown], label="Precision", multiselect=False, value="float16", interactive=True, ) weight_type = gr.Dropdown( choices=[i.value.name for i in WeightType], label="Weights type", multiselect=False, value="Original", interactive=True, ) base_model_name_textbox = gr.Textbox(label="Base model (for delta or adapter weights)") submit_button = gr.Button("Submit Eval") submission_result = gr.Markdown() submit_button.click( add_new_eval, [ model_name_textbox, base_model_name_textbox, revision_name_textbox, precision, weight_type, model_type, ], submission_result, ) with gr.Row(): with gr.Accordion("๐Ÿ“™ Citation", open=False): citation_button = gr.Textbox( value=CITATION_BUTTON_TEXT, label=CITATION_BUTTON_LABEL, lines=20, elem_id="citation-button", show_copy_button=True, ) scheduler = BackgroundScheduler() scheduler.add_job(restart_space, "interval", seconds=1800) scheduler.start() demo.queue(default_concurrency_limit=40).launch() # import gradio as gr # import pandas as pd # csv_filename = 'leaderboard.csv' # # url = 'https://docs.google.com/spreadsheets/d/1Oh3nrbdWjKuh9twJsc9yJLppiJeD_BZyKgCTOxRkALM/export?format=csv' # def get_data_classifica(): # dataset = pd.read_csv("leaderboard_general.csv", sep=',') # if 'model ' in dataset.columns: # dataset.rename(columns={'model ': 'model'}, inplace=True) # df_classifica = dataset[['model', 'helloswag_it acc norm', 'arc_it acc norm', 'm_mmlu_it acc shot 5']] # df_classifica['media'] = df_classifica[['helloswag_it acc norm', 'arc_it acc norm', 'm_mmlu_it acc shot 5']].mean(axis=1) # df_classifica['media'] = df_classifica['media'].round(3) # df_classifica = df_classifica.sort_values(by='media', ascending=False) # df_classifica = df_classifica[['model', 'media', 'helloswag_it acc norm', 'arc_it acc norm', 'm_mmlu_it acc shot 5']] # return df_classifica # def get_data_totale(): # dataset = pd.read_csv("leaderboard_general.csv", sep=',') # if 'model ' in dataset.columns: # dataset.rename(columns={'model ': 'model'}, inplace=True) # return dataset # with gr.Blocks() as demo: # with gr.Tab('Classifica Generale'): # gr.Markdown('''# Classifica generale degli LLM italiani''') # discord_link = 'https://discord.gg/m7sS3mduY2' # gr.Markdown(''' # I modelli sottostanti sono stati testati con [lm_evaluation_harness](https://github.com/EleutherAI/lm-evaluation-harness) su task specifici per l'italiano introdotti con questa [PR](https://github.com/EleutherAI/lm-evaluation-harness/pull/1358). # L'intero progetto, i modelli e i dataset sono rigorosamente open source e tutti i risultati sono riproducibili lanciando i seguenti comandi: # ``` # lm_eval --model hf --model_args pretrained=HUGGINGFACE_MODEL_ID --tasks hellaswag_it,arc_it --device cuda:0 --batch_size auto:2 # ``` # ``` # lm_eval --model hf --model_args pretrained=HUGGINGFACE_MODEL_ID --tasks m_mmlu_it --num_fewshot 5 --device cuda:0 --batch_size auto:2 # ``` # ''') # gr.DataFrame(get_data_classifica, every=3600) # gr.Markdown(f"Contributore principale: @giux78") # gr.Markdown(''' # ### Risultati su modelli "internazionali" (instruct) # | Model | Arc-c | HellaS | MMUL | AVG | # | --- | --- | --- | --- | --- | # | Mixtral 8x22b | 55.3 | 77.1 | 75.8 | 69.4 | # | LLama3 70b | 52.9 | 70.3 | 74.8 | 66.0 | # | command-r-plus | 49.5 | 74.9 | 67.6 | 64.0 | # | Mixtral 8x7b | 51.1 | 72.9 | 65.9 | 63.3 | # | LLama2 70b | 49.4 | 70.9 | 65.1 | 61.8 | # | command-r-v01 | 50.8 | 72.3 | 60.0 | 61.0 | # | Phi-3-mini | 43.46 | 61.44 | 56.55 | 53.8 | # | LLama3 8b | 44.3 | 59.9 | 55.7 | 53.3 | # | LLama1 34b | 42.9 | 65.4 | 49.0 | 52.4 | # | Mistral 7b | 41.49 | 61.22 | 52.53 | 51.7 | # | Gemma 1.1 7b | 41.75 | 54.07 | 49.45 | 48.4 | # ''') # with gr.Tab('Classifica RAG'): # gr.Markdown('''# Classifica RAG degli LLM italiani''') # gr.Markdown(f'''In questa sezione i modelli sono valutati su dei task di Q&A e ordinati per F1 Score e EM (Exact Match). La repo di riferimento รจ [questa](https://github.com/C080/open-llm-ita-leaderboard). # I modelli in cima alla classifica sono ritenuti preferibili per i task di Retrieval Augmented Generation.''') # gr.Dataframe(pd.read_csv(csv_filename, sep=';')) # gr.Markdown(f"Si ringrazia il @galatolo per il codice dell'eval.") # with gr.Tab('Eval aggiuntive'): # gr.Markdown('''# Altre evaluation''') # gr.Markdown('''Qui ci sono altri test di altri modelli, che non sono ancora stati integrati nella classifica generale.''') # gr.DataFrame(get_data_totale, every=3600) # with gr.Tab('Informazioni'): # form_link = "https://forms.gle/Gc9Dfu52xSBhQPpAA" # gr.Markdown('''# Community discord # Se vuoi contribuire al progetto o semplicemente unirti alla community di LLM italiani unisciti al nostro [discord!](https://discord.gg/m7sS3mduY2) # # Aggiungi il tuo modello # Se hai sviluppato un tuo modello che vuoi far valutare, compila il form [qui](https://forms.gle/Gc9Dfu52xSBhQPpAA) รจ tutto gratuito! # ''') # with gr.Tab('Sponsor'): # gr.Markdown(''' # # Sponsor # Le evaluation della classifica generale sono state gentilmente offerte da un provider cloud italiano [seeweb.it](https://www.seeweb.it/) specializzato in servizi di GPU cloud e AI. # ''') # demo.launch()