"""Script to produce radial plots.""" from functools import partial import plotly.graph_objects as go import json import numpy as np from collections import defaultdict import pandas as pd from pydantic import BaseModel import gradio as gr import requests import random import logging import datetime as dt import scipy.stats as stats import itertools as it fmt = "%(asctime)s [%(levelname)s] <%(name)s> %(message)s" logging.basicConfig(level=logging.INFO, format=fmt) logger = logging.getLogger("radial_plot_generator") INTRO_MARKDOWN = """ # Radial Plot Generator This demo allows you to generate a radial plot comparing the performance of different language models on different tasks. It is based on the generative results from the [ScandEval benchmark](https://scandeval.com). """ ABOUT_MARKDOWN = """ ## About the ScandEval Benchmark The [ScandEval benchmark](https://scandeval.com) is used compare pretrained language models on tasks in Danish, Swedish, Norwegian Bokmål, Norwegian Nynorsk, Icelandic, Faroese, German, Dutch and English. The benchmark supports both encoder models (such as BERT) and generative models (such as GPT), and leaderboards for both kinds [are available](https://scandeval.com). The generative models are evaluated using in-context learning with few-shot prompts. The few-shot examples are sampled randomly from the training split, and we benchmark the models 10 times with bootstrapped test sets and different few-shot examples in each iteration. This allows us to better measure the uncertainty of the results. We use the uncertainty in the radial plot when we compute the rank scores for the models. Namely, we compute the rank score by firstly computing the rank of the model on each task, where two models are considered to have the same rank if there is not a statistically significant difference between their scores (one-tailed t-test with p < 0.05). We next apply a logaritmic transformation to the ranks, to downplay the importance of the poorly performing models. Lastly, we invert and normalise the logaritmic ranks to the range [0, 1], resulting in the best performing models having rank scores close to 1 and the worst performing models having rank scores close to 0. ## The Benchmark Datasets The ScandEval generative benchmark currently covers the languages Danish, Swedish, Norwegian, Icelandic, German, Dutch and English. For each language, the benchmark consists of 7 different tasks, each of which consists of 1-2 datasets. The tasks are the following: ### Text Classification Given a piece of text, classify it into a number of classes. For this task we extract the first token of the possible labels, and choose the label whose first token has the highest probability. All datasets in this category are currently trinary sentiment classification datasets. We use the Matthews Correlation Coefficient (MCC) as the evaluation metric. ### Information Extraction Given a piece of text, extract a number of entities from the text. As the model needs to extract multiple entities, we use [structured generation](https://github.com/noamgat/lm-format-enforcer) to make the model generate a JSON dictionary with keys being the entity categories and values being lists of the identified entities. All datasets in this task are named entity recognition datasets. We use the micro-averaged F1 score as the evaluation metric, where we ignore the Miscellaneous category. ### Grammar Given a piece of text, determine whether it is grammatically correct or not. All datasets in this task are built from the dependency treebanks of the languages, where words are removed or swapped, in a way that makes the sentence ungrammatical. We use the Matthews Correlation Coefficient (MCC) as the evaluation metric. ### Question Answering Given a question and a piece of text, extract the answer to the question from the text. All datasets in this task are extractive question answering datasets. We use the exact match (EM) score as the evaluation metric. ### Summarisation Given a piece of text, generate a summary of the text. All the datasets come from either news articles or WikiHow articles. We use the BERTScore metric as the evaluation metric, where the encoder model used is [microsoft/mdeberta-v3-base](https://huggingface.co/microsoft/mdeberta-v3-base). ### Knowledge Given a trivia-style question with multiple choice answers, choose the correct answer. As with text classification, we use the probabilities of the answer letter (a, b, c or d) to choose the answer. The datasets in this task are machine translated versions of the [MMLU](https://doi.org/10.48550/arXiv.2009.03300) and [ARC](https://allenai.org/data/arc) datasets. We use the Matthews Correlation Coefficient (MCC) as the evaluation metric. ### Reasoning Given a scenario and multiple possible endings, choose the correct ending. As with text classification, we use the probabilities of the answer letter (a, b, c or d) to choose the answer. The datasets in this task are machine translated versions of the [HellaSwag](https://rowanzellers.com/hellaswag/) dataset. We use the Matthews Correlation Coefficient (MCC) as the evaluation metric. ## Citation If you use the ScandEval benchmark in your work, please cite [the paper](https://aclanthology.org/2023.nodalida-1.20): ``` @inproceedings{nielsen2023scandeval, title={ScandEval: A Benchmark for Scandinavian Natural Language Processing}, author={Nielsen, Dan}, booktitle={Proceedings of the 24th Nordic Conference on Computational Linguistics (NoDaLiDa)}, pages={185--201}, year={2023} } ``` """ UPDATE_FREQUENCY_MINUTES = 5 MIN_COLOUR_DISTANCE_BETWEEN_MODELS = 200 class Task(BaseModel): """Class to hold task information.""" name: str metric: str def __hash__(self): return hash(self.name) class Language(BaseModel): """Class to hold language information.""" code: str name: str def __hash__(self): return hash(self.code) class Dataset(BaseModel): """Class to hold dataset information.""" name: str language: Language task: Task def __hash__(self): return hash(self.name) SUMMARISATION = Task(name="summarisation", metric="bertscore") KNOWLEDGE = Task(name="knowledge", metric="mcc") REASONING = Task(name="reasoning", metric="mcc") GRAMMAR = Task(name="grammar", metric="mcc") QUESTION_ANSWERING = Task(name="question answering", metric="em") TEXT_CLASSIFICATION = Task(name="text classification", metric="mcc") INFORMATION_EXTRACTION = Task(name="information extraction", metric="micro_f1_no_misc") ALL_TASKS = [obj for obj in globals().values() if isinstance(obj, Task)] DANISH = Language(code="da", name="Danish") NORWEGIAN = Language(code="no", name="Norwegian") SWEDISH = Language(code="sv", name="Swedish") ICELANDIC = Language(code="is", name="Icelandic") GERMAN = Language(code="de", name="German") DUTCH = Language(code="nl", name="Dutch") ENGLISH = Language(code="en", name="English") ALL_LANGUAGES = { obj.name: obj for obj in globals().values() if isinstance(obj, Language) } DATASETS = [ Dataset(name="swerec", language=SWEDISH, task=TEXT_CLASSIFICATION), Dataset(name="angry-tweets", language=DANISH, task=TEXT_CLASSIFICATION), Dataset(name="norec", language=NORWEGIAN, task=TEXT_CLASSIFICATION), Dataset(name="sb10k", language=GERMAN, task=TEXT_CLASSIFICATION), Dataset(name="dutch-social", language=DUTCH, task=TEXT_CLASSIFICATION), Dataset(name="sst5", language=ENGLISH, task=TEXT_CLASSIFICATION), Dataset(name="suc3", language=SWEDISH, task=INFORMATION_EXTRACTION), Dataset(name="dansk", language=DANISH, task=INFORMATION_EXTRACTION), Dataset(name="norne-nb", language=NORWEGIAN, task=INFORMATION_EXTRACTION), Dataset(name="norne-nn", language=NORWEGIAN, task=INFORMATION_EXTRACTION), Dataset(name="mim-gold-ner", language=ICELANDIC, task=INFORMATION_EXTRACTION), Dataset(name="germeval", language=GERMAN, task=INFORMATION_EXTRACTION), Dataset(name="conll-nl", language=DUTCH, task=INFORMATION_EXTRACTION), Dataset(name="conll-en", language=ENGLISH, task=INFORMATION_EXTRACTION), Dataset(name="scala-sv", language=SWEDISH, task=GRAMMAR), Dataset(name="scala-da", language=DANISH, task=GRAMMAR), Dataset(name="scala-nb", language=NORWEGIAN, task=GRAMMAR), Dataset(name="scala-nn", language=NORWEGIAN, task=GRAMMAR), Dataset(name="scala-is", language=ICELANDIC, task=GRAMMAR), Dataset(name="scala-de", language=GERMAN, task=GRAMMAR), Dataset(name="scala-nl", language=DUTCH, task=GRAMMAR), Dataset(name="scala-en", language=ENGLISH, task=GRAMMAR), Dataset(name="scandiqa-da", language=DANISH, task=QUESTION_ANSWERING), Dataset(name="norquad", language=NORWEGIAN, task=QUESTION_ANSWERING), Dataset(name="scandiqa-sv", language=SWEDISH, task=QUESTION_ANSWERING), Dataset(name="nqii", language=ICELANDIC, task=QUESTION_ANSWERING), Dataset(name="germanquad", language=GERMAN, task=QUESTION_ANSWERING), Dataset(name="squad", language=ENGLISH, task=QUESTION_ANSWERING), Dataset(name="squad-nl", language=DUTCH, task=QUESTION_ANSWERING), Dataset(name="nordjylland-news", language=DANISH, task=SUMMARISATION), Dataset(name="mlsum", language=GERMAN, task=SUMMARISATION), Dataset(name="rrn", language=ICELANDIC, task=SUMMARISATION), Dataset(name="no-sammendrag", language=NORWEGIAN, task=SUMMARISATION), Dataset(name="wiki-lingua-nl", language=DUTCH, task=SUMMARISATION), Dataset(name="swedn", language=SWEDISH, task=SUMMARISATION), Dataset(name="cnn-dailymail", language=ENGLISH, task=SUMMARISATION), Dataset(name="danish-citizen-tests", language=DANISH, task=KNOWLEDGE), Dataset(name="danske-talemaader", language=DANISH, task=KNOWLEDGE), Dataset(name="mmlu-no", language=NORWEGIAN, task=KNOWLEDGE), Dataset(name="mmlu-sv", language=SWEDISH, task=KNOWLEDGE), Dataset(name="mmlu-is", language=ICELANDIC, task=KNOWLEDGE), Dataset(name="mmlu-de", language=GERMAN, task=KNOWLEDGE), Dataset(name="mmlu-nl", language=DUTCH, task=KNOWLEDGE), Dataset(name="mmlu", language=ENGLISH, task=KNOWLEDGE), Dataset(name="hellaswag-da", language=DANISH, task=REASONING), Dataset(name="hellaswag-no", language=NORWEGIAN, task=REASONING), Dataset(name="hellaswag-sv", language=SWEDISH, task=REASONING), Dataset(name="hellaswag-is", language=ICELANDIC, task=REASONING), Dataset(name="hellaswag-de", language=GERMAN, task=REASONING), Dataset(name="hellaswag-nl", language=DUTCH, task=REASONING), Dataset(name="hellaswag", language=ENGLISH, task=REASONING), ] def main() -> None: """Produce a radial plot.""" global last_fetch results_dfs = fetch_results() last_fetch = dt.datetime.now() all_languages = sorted( [language.name for language in ALL_LANGUAGES.values()], key=lambda language_name: language_name.lower(), ) danish_models = sorted( list({model_id for model_id in results_dfs[DANISH].index}), key=lambda model_id: model_id.lower(), ) global colour_mapping global seed seed = 4242 update_colour_mapping(results_dfs=results_dfs) with gr.Blocks(theme=gr.themes.Monochrome()) as demo: gr.Markdown(INTRO_MARKDOWN) with gr.Tab(label="Build a Radial Plot"): with gr.Column(): with gr.Row(): language_names_dropdown = gr.Dropdown( choices=all_languages, multiselect=True, label="Languages", value=["Danish"], interactive=True, scale=2, ) model_ids_dropdown = gr.Dropdown( choices=danish_models, multiselect=True, label="Models", value=["gpt-4-0613", "mistralai/Mistral-7B-v0.1"], interactive=True, scale=2, ) with gr.Row(): use_rank_score_checkbox = gr.Checkbox( label="Compare models with rank scores (as opposed to raw " "scores)", value=True, interactive=True, scale=1, ) show_scale_checkbox = gr.Checkbox( label="Show the scale on the plot (always 0-100)", value=False, interactive=True, scale=1, ) plot_width_slider = gr.Slider( label="Plot width", minimum=600, maximum=1000, step=10, value=800, interactive=True, scale=1, ) plot_height_slider = gr.Slider( label="Plot height", minimum=300, maximum=700, step=10, value=500, interactive=True, scale=1, ) update_colours_button = gr.Button( value="Update colours", interactive=True, scale=1, ) with gr.Row(): plot = gr.Plot( value=produce_radial_plot( model_ids_dropdown.value, language_names=language_names_dropdown.value, use_rank_score=use_rank_score_checkbox.value, show_scale=show_scale_checkbox.value, plot_width=plot_width_slider.value, plot_height=plot_height_slider.value, results_dfs=results_dfs, ), ) with gr.Tab(label="About"): gr.Markdown(ABOUT_MARKDOWN) gr.Markdown( "
Made with ❤️ by the " "Alexandra Institute.
" ) language_names_dropdown.change( fn=partial(update_model_ids_dropdown, results_dfs=results_dfs), inputs=[language_names_dropdown, model_ids_dropdown], outputs=model_ids_dropdown, ) # Update plot when anything changes update_plot_kwargs = dict( fn=partial( produce_radial_plot, results_dfs=results_dfs, ), inputs=[ model_ids_dropdown, language_names_dropdown, use_rank_score_checkbox, show_scale_checkbox, plot_width_slider, plot_height_slider, ], outputs=plot, ) language_names_dropdown.change(**update_plot_kwargs) model_ids_dropdown.change(**update_plot_kwargs) use_rank_score_checkbox.change(**update_plot_kwargs) show_scale_checkbox.change(**update_plot_kwargs) plot_width_slider.change(**update_plot_kwargs) plot_height_slider.change(**update_plot_kwargs) # Update colours when the button is clicked update_colours_button.click( fn=partial(update_colour_mapping, results_dfs=results_dfs), ).then(**update_plot_kwargs) demo.launch() def update_model_ids_dropdown( language_names: list[str], model_ids: list[str], results_dfs: dict[Language, pd.DataFrame] | None, ) -> dict: """When the language names are updated, update the model ids dropdown. Args: language_names: The names of the languages to include in the plot. model_ids: The ids of the models to include in the plot. results_dfs: The results dataframes for each language. Returns: The Gradio update to the model ids dropdown. """ global last_fetch minutes_since_last_fetch = (dt.datetime.now() - last_fetch).total_seconds() / 60 if minutes_since_last_fetch > UPDATE_FREQUENCY_MINUTES: results_dfs = fetch_results() last_fetch = dt.datetime.now() if results_dfs is None or len(language_names) == 0: if results_dfs is None: logger.info("No results fetched yet. Resetting model ids dropdown.") else: logger.info("No languages selected. Resetting model ids dropdown.") return gr.update(choices=[], value=[]) tasks = [ task for task in ALL_TASKS if all( task in df.columns for language, df in results_dfs.items() if language.name in language_names ) ] filtered_results_dfs = { language: df[tasks] for language, df in results_dfs.items() if language.name in language_names } unique_models: set[str] = { str(model_id) for df in filtered_results_dfs.values() for model_id in df.index } filtered_models: list[str] = sorted( [ model_id for model_id in unique_models if all(model_id in df.index for df in filtered_results_dfs.values()) ], key=lambda model_id: model_id.lower(), ) if len(filtered_models) == 0: logger.info( "No valid models for the selected languages. Resetting model ids dropdown." ) return gr.update(choices=[], value=[]) valid_selected_models: list[str] = [ model_id for model_id in model_ids if model_id in filtered_models ] if not valid_selected_models: if len(filtered_models) > 1: valid_selected_models = random.sample(population=filtered_models, k=2) elif len(filtered_models) == 1: valid_selected_models = random.sample(population=filtered_models, k=1) logger.info( f"Updated model ids dropdown with {len(filtered_models):,} valid models for " f"the selected languages, with {valid_selected_models} selected." ) return gr.update(choices=filtered_models, value=valid_selected_models) def produce_radial_plot( model_ids: list[str], language_names: list[str], use_rank_score: bool, show_scale: bool, plot_width: int, plot_height: int, results_dfs: dict[Language, pd.DataFrame] | None, ) -> go.Figure: """Produce a radial plot as a plotly figure. Args: model_ids: The ids of the models to include in the plot. language_names: The names of the languages to include in the plot. use_rank_score: Whether to use rank scores (as opposed to raw scores). show_scale: Whether to show the scale on the plot. plot_width: The width of the plot. plot_height: The height of the plot. results_dfs: The results dataframes for each language. Returns: A plotly figure. """ global last_fetch minutes_since_last_fetch = (dt.datetime.now() - last_fetch).total_seconds() / 60 if minutes_since_last_fetch > UPDATE_FREQUENCY_MINUTES: results_dfs = fetch_results() last_fetch = dt.datetime.now() if results_dfs is None or len(language_names) == 0 or len(model_ids) == 0: if results_dfs is None: logger.info("No results fetched yet. Resetting plot.") elif len(language_names) == 0: logger.info("No languages selected. Resetting plot.") else: logger.info("No models selected. Resetting plot.") return go.Figure() logger.info( f"Producing radial plot for models {model_ids!r} on languages " f"{language_names!r}..." ) languages = [ALL_LANGUAGES[language_name] for language_name in language_names] results_dfs_filtered = { language: df for language, df in results_dfs.items() if language.name in language_names } tasks = [ task for task in ALL_TASKS if all(task in df.columns for df in results_dfs_filtered.values()) ] logger.info("Computing rank scores...") all_rank_scores: dict[Task, dict[Language, dict[str, float]]] = { task: { language: dict() for language in languages } for task in tasks } for task in tasks: for language in languages: df = results_dfs_filtered[language][task].dropna() model_ids_sorted: list[str] = ( df.map(np.mean).sort_values(ascending=False).index.tolist() ) ranks = list() rank = 0 best_scores = None for model_id in model_ids_sorted: if best_scores is None: best_scores = df.loc[model_id] rank = 1 else: scores = df.loc[model_id] worse_than_previous_models = stats.ttest_rel( a=best_scores, b=scores, alternative="greater" ).pvalue < 0.05 if worse_than_previous_models: rank += 1 best_scores = scores ranks.append(rank) log_ranks = np.log(ranks) scores = 1 - (log_ranks / log_ranks.max()) for model_id, score in zip(model_ids_sorted, scores): all_rank_scores[task][language][model_id] = score logger.info("Successfully computed rank scores.") # Add all the evaluation results for each model results: list[list[float]] = list() for model_id in model_ids: result_list = list() for task in tasks: rank_scores = list() scores = list() for language in languages: if model_id not in results_dfs_filtered[language].index: continue rank_score = 100 * all_rank_scores[task][language][model_id] rank_scores.append(rank_score) score_arr = np.array(results_dfs_filtered[language].loc[model_id][task]) if score_arr.mean() < 1: score_arr *= 100 scores.append(score_arr.mean()) if use_rank_score: result_list.append(np.mean(rank_scores)) else: result_list.append(np.mean(scores)) results.append(result_list) # Get a matrix of shape [num_models, num_tasks], where entry (i, j) indicates how # many models that model i has beaten on task j result_matrix = np.array(results) num_models = result_matrix.shape[0] num_tasks = result_matrix.shape[1] num_models_beaten = np.zeros((num_models, num_tasks)) for i in range(num_models): for j in range(num_tasks): num_models_beaten[i, j] = np.sum( result_matrix[i, j] > result_matrix[:, j] ) # Sort the models (and their results) such that the model who beats most other # models first. This will result in the "smaller areas" being on top of the "larger # areas", which is more aesthetically pleasing. sorted_idxs = num_models_beaten.sum(axis=1).argsort()[::-1] model_ids = np.asarray(model_ids)[sorted_idxs].tolist() results = result_matrix[sorted_idxs].tolist() # Add the results to a plotly figure fig = go.Figure() for model_id, result_list in zip(model_ids, results): r, g, b = colour_mapping[model_id] fig.add_trace(go.Scatterpolar( r=result_list, theta=[task.name for task in tasks], name=model_id, fill='toself', fillcolor=f'rgba({r}, {g}, {b}, 0.6)', line=dict(color=f'rgb({r}, {g}, {b})'), )) languages_str = "" if len(languages) > 1: languages_str = ", ".join([language.name for language in languages[:-1]]) languages_str += " and " languages_str += languages[-1].name if use_rank_score: title = f'Rank Score on on {languages_str} Language Tasks' else: title = f'Raw Score on on {languages_str} Language Tasks' # Builds the radial plot from the results fig.update_layout( polar=dict(radialaxis=dict(visible=show_scale, range=[0, 100])), showlegend=True, title=title, width=plot_width, height=plot_height, ) logger.info("Successfully produced radial plot.") return fig def fetch_results() -> dict[Language, pd.DataFrame]: """Fetch the results from the ScandEval benchmark. Returns: A dictionary of languages -> results-dataframes, whose indices are the models and columns are the tasks. """ logger.info("Fetching results from ScandEval benchmark...") response = requests.get( "https://www.scandeval.com/scandeval_benchmark_results.jsonl" ) response.raise_for_status() records = [ json.loads(dct_str) for dct_str in response.text.split("\n") if dct_str.strip("\n") ] # Build a dictionary of languages -> results-dataframes, whose indices are the # models and columns are the tasks. results_dfs = dict() for language in {dataset.language for dataset in DATASETS}: possible_dataset_names = { dataset.name for dataset in DATASETS if dataset.language == language } data_dict = defaultdict(dict) for record in records: model_name = record["model"] # Manual fix for OpenAI models: Only keep the validation split results if "gpt-3.5" in model_name or "gpt-4" in model_name: if not record.get("validation_split", False): continue dataset_name = record["dataset"] if dataset_name in possible_dataset_names: dataset = next( dataset for dataset in DATASETS if dataset.name == dataset_name ) scores = [ test_score_dict.get( f"test_{dataset.task.metric}", test_score_dict.get(dataset.task.metric) ) for test_score_dict in record["results"]["raw"]["test"] ] if dataset.task in data_dict[model_name]: data_dict[model_name][dataset.task].append(scores) else: data_dict[model_name][dataset.task] = [scores] results_df = pd.DataFrame(data_dict).T.map( lambda lists_or_nan: list(it.chain(lists_or_nan)) if lists_or_nan == lists_or_nan else lists_or_nan ).dropna().map(lambda lst: lst[0]) results_dfs[language] = results_df logger.info("Successfully fetched results from ScandEval benchmark.") return results_dfs def update_colour_mapping(results_dfs: dict[Language, pd.DataFrame]) -> None: """Get a mapping from model ids to RGB triplets. Args: results_dfs: The results dataframes for each language. """ global colour_mapping global seed seed += 1 gr.Info(f"Updating colour mapping...") # Get distinct RGB values for all models all_models = list( {model_id for df in results_dfs.values() for model_id in df.index} ) colour_mapping = dict() for i in it.count(): min_colour_distance = MIN_COLOUR_DISTANCE_BETWEEN_MODELS - i retries_left = 10 * len(all_models) for model_id in all_models: random.seed(hash(model_id) + i + seed) r, g, b = 0, 0, 0 too_bright, similar_to_other_model = True, True while (too_bright or similar_to_other_model) and retries_left > 0: r, g, b = tuple(random.randint(0, 255) for _ in range(3)) too_bright = np.min([r, g, b]) > 200 similar_to_other_model = any( np.abs( np.array(colour) - np.array([r, g, b]) ).sum() < min_colour_distance for colour in colour_mapping.values() ) retries_left -= 1 colour_mapping[model_id] = (r, g, b) if retries_left: logger.info( f"Successfully found a colour mapping with min colour distance " f"{min_colour_distance}." ) break if __name__ == "__main__": main()