import argparse import textwrap from multiprocessing import Manager, Pool import pandas as pd import plotly.express as px import streamlit as st from datasets import get_dataset_infos from pygments import highlight from pygments.formatters import HtmlFormatter from pygments.lexers import DjangoLexer from promptsource.session import _get_state from promptsource.templates import Template, TemplateCollection from promptsource.utils import ( get_dataset, get_dataset_confs, list_datasets, removeHyphen, renameDatasetColumn, render_features, ) # add an argument for read-only # At the moment, streamlit does not handle python script arguments gracefully. # Thus, for read-only mode, you have to type one of the below two: # streamlit run promptsource/app.py -- -r # streamlit run promptsource/app.py -- --read-only # Check https://github.com/streamlit/streamlit/issues/337 for more information. parser = argparse.ArgumentParser(description="run app.py with args") parser.add_argument("-r", "--read-only", action="store_true", help="whether to run it as read-only mode") args = parser.parse_args() if args.read_only: select_options = ["Helicopter view", "Prompted dataset viewer"] side_bar_title_prefix = "Promptsource (Read only)" else: select_options = ["Helicopter view", "Prompted dataset viewer", "Sourcing"] side_bar_title_prefix = "Promptsource" # # Helper functions for datasets library # get_dataset = st.cache(allow_output_mutation=True)(get_dataset) get_dataset_confs = st.cache(get_dataset_confs) def reset_template_state(): state.template_name = None state.jinja = None state.reference = None # # Loads session state # state = _get_state() # # Initial page setup # st.set_page_config(page_title="Promptsource", layout="wide") st.sidebar.markdown( "
πŸ’»Github - Promptsource\n\n
", unsafe_allow_html=True, ) mode = st.sidebar.selectbox( label="Choose a mode", options=select_options, index=0, key="mode_select", ) st.sidebar.title(f"{side_bar_title_prefix} 🌸 - {mode}") # # Adds pygments styles to the page. # st.markdown( "", unsafe_allow_html=True ) WIDTH = 80 def show_jinja(t, width=WIDTH): wrap = textwrap.fill(t, width=width, replace_whitespace=False) out = highlight(wrap, DjangoLexer(), HtmlFormatter()) st.write(out, unsafe_allow_html=True) def show_text(t, width=WIDTH, with_markdown=False): wrap = [textwrap.fill(subt, width=width, replace_whitespace=False) for subt in t.split("\n")] wrap = "\n".join(wrap) if with_markdown: st.write(wrap, unsafe_allow_html=True) else: st.text(wrap) # # Loads template data # try: template_collection = TemplateCollection() except FileNotFoundError: st.error( "Unable to find the prompt folder!\n\n" "We expect the folder to be in the working directory. " "You might need to restart the app in the root directory of the repo." ) st.stop() if mode == "Helicopter view": st.title("High level metrics") st.write( "If you want to contribute, please refer to the instructions in " + "[Contributing](https://github.com/bigscience-workshop/promptsource/blob/main/CONTRIBUTING.md)." ) # # Global metrics # counts = template_collection.get_templates_count() nb_prompted_datasets = len(counts) st.write(f"## Number of *prompted datasets*: `{nb_prompted_datasets}`") nb_prompts = sum(counts.values()) st.write(f"## Number of *prompts*: `{nb_prompts}`") # # Metrics per dataset/subset # # Download dataset infos (multiprocessing download) manager = Manager() all_infos = manager.dict() all_datasets = list(set([t[0] for t in template_collection.keys])) def get_infos(d_name): all_infos[d_name] = get_dataset_infos(d_name) pool = Pool(processes=len(all_datasets)) pool.map(get_infos, all_datasets) pool.close() pool.join() results = [] for (dataset_name, subset_name) in template_collection.keys: # Collect split sizes (train, validation and test) if dataset_name not in all_infos: infos = get_dataset_infos(dataset_name) all_infos[dataset_name] = infos else: infos = all_infos[dataset_name] if infos: if subset_name is None: subset_infos = infos[list(infos.keys())[0]] else: subset_infos = infos[subset_name] split_sizes = {k: v.num_examples for k, v in subset_infos.splits.items()} else: # Zaid/coqa_expanded and Zaid/quac_expanded don't have dataset_infos.json # so infos is an empty dic, and `infos[list(infos.keys())[0]]` raises an error # For simplicity, just filling `split_sizes` with nothing, so the displayed split sizes will be 0. split_sizes = {} # Collect template counts, original task counts and names dataset_templates = template_collection.get_dataset(dataset_name, subset_name) results.append( { "Dataset name": dataset_name, "Subset name": "βˆ…" if subset_name is None else subset_name, "Train size": split_sizes["train"] if "train" in split_sizes else 0, "Validation size": split_sizes["validation"] if "validation" in split_sizes else 0, "Test size": split_sizes["test"] if "test" in split_sizes else 0, "Number of prompts": len(dataset_templates), "Number of original task prompts": sum( [bool(t.metadata.original_task) for t in dataset_templates.templates.values()] ), "Prompt names": [t.name for t in dataset_templates.templates.values()], } ) results_df = pd.DataFrame(results) results_df.sort_values(["Number of prompts"], inplace=True, ascending=False) results_df.reset_index(drop=True, inplace=True) nb_training_instances = results_df["Train size"].sum() st.write(f"## Number of *training instances*: `{nb_training_instances}`") plot_df = results_df[["Dataset name", "Subset name", "Train size", "Number of prompts"]].copy() plot_df["Name"] = plot_df["Dataset name"] + " - " + plot_df["Subset name"] plot_df.sort_values(["Train size"], inplace=True, ascending=False) fig = px.bar( plot_df, x="Name", y="Train size", hover_data=["Dataset name", "Subset name", "Number of prompts"], log_y=True, title="Number of training instances per data(sub)set - y-axis is in logscale", ) fig.update_xaxes(visible=False, showticklabels=False) st.plotly_chart(fig, use_container_width=True) st.write( f"- Top 3 training subsets account for `{100*plot_df[:3]['Train size'].sum()/nb_training_instances:.2f}%` of the training instances." ) biggest_training_subset = plot_df.iloc[0] st.write( f"- Biggest training subset is *{biggest_training_subset['Name']}* with `{biggest_training_subset['Train size']}` instances" ) smallest_training_subset = plot_df[plot_df["Train size"] > 0].iloc[-1] st.write( f"- Smallest training subset is *{smallest_training_subset['Name']}* with `{smallest_training_subset['Train size']}` instances" ) st.markdown("***") st.write("Details per dataset") st.table(results_df) else: # Combining mode `Prompted dataset viewer` and `Sourcing` since the # backbone of the interfaces is the same assert mode in ["Prompted dataset viewer", "Sourcing"], ValueError( f"`mode` ({mode}) should be in `[Helicopter view, Prompted dataset viewer, Sourcing]`" ) # # Loads dataset information # dataset_list = list_datasets( template_collection, state, ) ag_news_index = dataset_list.index("ag_news") # # Select a dataset - starts with ag_news # dataset_key = st.sidebar.selectbox( "Dataset", dataset_list, key="dataset_select", index=ag_news_index, help="Select the dataset to work on.", ) # # If a particular dataset is selected, loads dataset and template information # if dataset_key is not None: # # Check for subconfigurations (i.e. subsets) # configs = get_dataset_confs(dataset_key) conf_option = None if len(configs) > 0: conf_option = st.sidebar.selectbox("Subset", configs, index=0, format_func=lambda a: a.name) dataset = get_dataset(dataset_key, str(conf_option.name) if conf_option else None) splits = list(dataset.keys()) index = 0 if "train" in splits: index = splits.index("train") split = st.sidebar.selectbox("Split", splits, key="split_select", index=index) dataset = dataset[split] dataset = renameDatasetColumn(dataset) dataset_templates = template_collection.get_dataset(dataset_key, conf_option.name if conf_option else None) template_list = dataset_templates.all_template_names num_templates = len(template_list) st.sidebar.write( "No of prompts created for " + f"`{dataset_key + (('/' + conf_option.name) if conf_option else '')}`" + f": **{str(num_templates)}**" ) if mode == "Prompted dataset viewer": if num_templates > 0: template_name = st.sidebar.selectbox( "Prompt name", template_list, key="template_select", index=0, help="Select the prompt to visualize.", ) step = 50 example_index = st.sidebar.number_input( f"Select the example index (Size = {len(dataset)})", min_value=0, max_value=len(dataset) - step, value=0, step=step, key="example_index_number_input", help="Offset = 50.", ) else: # mode = Sourcing st.sidebar.subheader("Select Example") example_index = st.sidebar.slider("Select the example index", 0, len(dataset) - 1) example = dataset[example_index] example = removeHyphen(example) st.sidebar.write(example) st.sidebar.subheader("Dataset Schema") rendered_features = render_features(dataset.features) st.sidebar.write(rendered_features) # # Display dataset information # st.header("Dataset: " + dataset_key + " " + (("/ " + conf_option.name) if conf_option else "")) st.markdown( "*Homepage*: " + dataset.info.homepage + "\n\n*Dataset*: https://github.com/huggingface/datasets/blob/master/datasets/%s/%s.py" % (dataset_key, dataset_key) ) md = """ %s """ % ( dataset.info.description.replace("\\", "") if dataset_key else "" ) st.markdown(md) # # Body of the app: display prompted examples in mode `Prompted dataset viewer` # or text boxes to create new prompts in mode `Sourcing` # if mode == "Prompted dataset viewer": # # Display template information # if num_templates > 0: template = dataset_templates[template_name] st.subheader("Prompt") st.markdown("##### Name") st.text(template.name) st.markdown("##### Reference") st.text(template.reference) st.markdown("##### Original Task? ") st.text(template.metadata.original_task) st.markdown("##### Choices in template? ") st.text(template.metadata.choices_in_prompt) st.markdown("##### Metrics") st.text(", ".join(template.metadata.metrics) if template.metadata.metrics else None) st.markdown("##### Answer Choices") if template.get_answer_choices_expr() is not None: show_jinja(template.get_answer_choices_expr()) else: st.text(None) st.markdown("##### Jinja template") splitted_template = template.jinja.split("|||") st.markdown("###### Input template") show_jinja(splitted_template[0].strip()) if len(splitted_template) > 1: st.markdown("###### Target template") show_jinja(splitted_template[1].strip()) st.markdown("***") # # Display a couple (steps) examples # for ex_idx in range(example_index, example_index + step): if ex_idx >= len(dataset): continue example = dataset[ex_idx] example = removeHyphen(example) col1, _, col2 = st.beta_columns([12, 1, 12]) with col1: st.write(example) if num_templates > 0: with col2: prompt = template.apply(example, highlight_variables=False) if prompt == [""]: st.write("βˆ…βˆ…βˆ… *Blank result*") else: st.write("Input") show_text(prompt[0]) if len(prompt) > 1: st.write("Target") show_text(prompt[1]) st.markdown("***") else: # mode = Sourcing st.markdown("## Prompt Creator") # # Create a new template or select an existing one # col1a, col1b, _, col2 = st.beta_columns([9, 9, 1, 6]) # current_templates_key and state.templates_key are keys for the templates object current_templates_key = (dataset_key, conf_option.name if conf_option else None) # Resets state if there has been a change in templates_key if state.templates_key != current_templates_key: state.templates_key = current_templates_key reset_template_state() with col1a, st.form("new_template_form"): new_template_name = st.text_input( "Create a New Prompt", key="new_template", value="", help="Enter name and hit enter to create a new prompt.", ) new_template_submitted = st.form_submit_button("Create") if new_template_submitted: if new_template_name in dataset_templates.all_template_names: st.error( f"A prompt with the name {new_template_name} already exists " f"for dataset {state.templates_key}." ) elif new_template_name == "": st.error("Need to provide a prompt name.") else: template = Template(new_template_name, "", "") dataset_templates.add_template(template) reset_template_state() state.template_name = new_template_name else: state.new_template_name = None with col1b, st.beta_expander("or Select Prompt", expanded=True): dataset_templates = template_collection.get_dataset(*state.templates_key) template_list = dataset_templates.all_template_names if state.template_name: index = template_list.index(state.template_name) else: index = 0 state.template_name = st.selectbox( "", template_list, key="template_select", index=index, help="Select the prompt to work on." ) if st.button("Delete Prompt", key="delete_prompt"): dataset_templates.remove_template(state.template_name) reset_template_state() variety_guideline = """ :heavy_exclamation_mark::question:Creating a diverse set of prompts whose differences go beyond surface wordings (i.e. marginally changing 2 or 3 words) is highly encouraged. Ultimately, the hope is that exposing the model to such a diversity will have a non-trivial impact on the model's robustness to the prompt formulation. \r**To get various prompts, you can try moving the cursor along theses axes**: \n- **Interrogative vs affirmative form**: Ask a question about an attribute of the inputs or tell the model to decide something about the input. \n- **Task description localization**: where is the task description blended with the inputs? In the beginning, in the middle, at the end? \n- **Implicit situation or contextualization**: how explicit is the query? For instance, *Given this review, would you buy this product?* is an indirect way to ask whether the review is positive. """ col1, _, _ = st.beta_columns([18, 1, 6]) with col1: if state.template_name is not None: show_text(variety_guideline, with_markdown=True) # # Edit the created or selected template # col1, _, col2 = st.beta_columns([18, 1, 6]) with col1: if state.template_name is not None: template = dataset_templates[state.template_name] # # If template is selected, displays template editor # with st.form("edit_template_form"): updated_template_name = st.text_input("Name", value=template.name) state.reference = st.text_input( "Prompt Reference", help="Short description of the prompt and/or paper reference for the prompt.", value=template.reference, ) # Metadata state.metadata = template.metadata state.metadata.original_task = st.checkbox( "Original Task?", value=template.metadata.original_task, help="Prompt asks model to perform the original task designed for this dataset.", ) state.metadata.choices_in_prompt = st.checkbox( "Choices in Template?", value=template.metadata.choices_in_prompt, help="Prompt explicitly lists choices in the template for the output.", ) # Metrics from here: # https://github.com/google-research/text-to-text-transfer-transformer/blob/4b580f23968c2139be7fb1cd53b22c7a7f686cdf/t5/evaluation/metrics.py metrics_choices = [ "BLEU", "ROUGE", "Squad", "Trivia QA", "Accuracy", "Pearson Correlation", "Spearman Correlation", "MultiRC", "AUC", "COQA F1", "Edit Distance", ] # Add mean reciprocal rank metrics_choices.append("Mean Reciprocal Rank") # Add generic other metrics_choices.append("Other") # Sort alphabetically metrics_choices = sorted(metrics_choices) state.metadata.metrics = st.multiselect( "Metrics", metrics_choices, default=template.metadata.metrics, help="Select all metrics that are commonly used (or should " "be used if a new task) to evaluate this prompt.", ) # Answer choices if template.get_answer_choices_expr() is not None: answer_choices = template.get_answer_choices_expr() else: answer_choices = "" state.answer_choices = st.text_input( "Answer Choices", value=answer_choices, help="A Jinja expression for computing answer choices. " "Separate choices with a triple bar (|||).", ) # Jinja state.jinja = st.text_area("Template", height=40, value=template.jinja) # Submit form if st.form_submit_button("Save"): if ( updated_template_name in dataset_templates.all_template_names and updated_template_name != state.template_name ): st.error( f"A prompt with the name {updated_template_name} already exists " f"for dataset {state.templates_key}." ) elif updated_template_name == "": st.error("Need to provide a prompt name.") else: # Parses state.answer_choices if state.answer_choices == "": updated_answer_choices = None else: updated_answer_choices = state.answer_choices dataset_templates.update_template( state.template_name, updated_template_name, state.jinja, state.reference, state.metadata, updated_answer_choices, ) # Update the state as well state.template_name = updated_template_name # # Displays template output on current example if a template is selected # (in second column) # with col2: if state.template_name is not None: st.empty() template = dataset_templates[state.template_name] prompt = template.apply(example) if prompt == [""]: st.write("βˆ…βˆ…βˆ… *Blank result*") else: st.write("Input") show_text(prompt[0], width=40) if len(prompt) > 1: st.write("Target") show_text(prompt[1], width=40) # # Must sync state at end # state.sync()