import gradio as gr import numpy as np import json import pandas as pd from openai import OpenAI import yaml from typing import Optional, List, Dict, Tuple, Any from topk_sae import FastAutoencoder import torch import plotly.express as px from collections import Counter from huggingface_hub import hf_hub_download import os import os print(os.getenv('MODEL_REPO_ID')) # Constants EMBEDDING_MODEL = "text-embedding-3-small" d_model = 1536 n_dirs = d_model * 6 k = 64 auxk = 128 device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') torch.set_grad_enabled(False) # Function to download all necessary files def download_all_files(): files_to_download = [ "astroPH_paper_metadata.csv", "csLG_feature_analysis_results_64.json", "astroPH_topk_indices_64_9216_int32.npy", "astroPH_64_9216.pth", "astroPH_topk_values_64_9216_float16.npy", "csLG_abstract_texts.json", "csLG_topk_values_64_9216_float16.npy", "csLG_abstract_embeddings_float16.npy", "csLG_paper_metadata.csv", "csLG_64_9216.pth", "astroPH_abstract_texts.json", "astroPH_feature_analysis_results_64.json", "csLG_topk_indices_64_9216_int32.npy", "astroPH_abstract_embeddings_float16.npy", # "csLG_clean_families_64_9216.json", # "astroPH_clean_families_64_9216.json", "astroPH_family_analysis_64_9216.json", "csLG_family_analysis_64_9216.json" ] for file in files_to_download: local_path = os.path.join("data", file) os.makedirs(os.path.dirname(local_path), exist_ok=True) hf_hub_download(repo_id="charlieoneill/saerch-ai-data", filename=file, local_dir="data") print(f"Downloaded {file}") # Load configuration and initialize OpenAI client download_all_files() # Load the API key from the environment variable api_key = os.getenv('openai_key') # Ensure the API key is set if not api_key: raise ValueError("The environment variable 'openai_key' is not set.") # Initialize the OpenAI client with the API key client = OpenAI(api_key=api_key) # Function to load data for a specific subject def load_subject_data(subject): embeddings_path = f"data/{subject}_abstract_embeddings_float16.npy" texts_path = f"data/{subject}_abstract_texts.json" feature_analysis_path = f"data/{subject}_feature_analysis_results_{k}.json" metadata_path = f'data/{subject}_paper_metadata.csv' topk_indices_path = f"data/{subject}_topk_indices_{k}_{n_dirs}_int32.npy" topk_values_path = f"data/{subject}_topk_values_{k}_{n_dirs}_float16.npy" families_path = f"data/{subject}_clean_families_{k}_{n_dirs}.json" family_analysis_path = f"data/{subject}_family_analysis_{k}_{n_dirs}.json" abstract_embeddings = np.load(embeddings_path).astype(np.float32) # Load float16 and convert to float32 with open(texts_path, 'r') as f: abstract_texts = json.load(f) with open(feature_analysis_path, 'r') as f: feature_analysis = json.load(f) df_metadata = pd.read_csv(metadata_path) topk_indices = np.load(topk_indices_path) # Already in int32, no conversion needed topk_values = np.load(topk_values_path).astype(np.float32) model_filename = f"{subject}_64_9216.pth" model_path = os.path.join("data", model_filename) ae = FastAutoencoder(n_dirs, d_model, k, auxk, multik=0).to(device) ae.load_state_dict(torch.load(model_path)) ae.eval() weights = torch.load(model_path) decoder = weights['decoder.weight'].cpu().numpy() del weights with open(family_analysis_path, 'r') as f: family_analysis = json.load(f) return { 'abstract_embeddings': abstract_embeddings, 'abstract_texts': abstract_texts, 'feature_analysis': feature_analysis, 'df_metadata': df_metadata, 'topk_indices': topk_indices, 'topk_values': topk_values, 'ae': ae, 'decoder': decoder, # 'feature_families': feature_families, 'family_analysis': family_analysis } # Load data for both subjects subject_data = { 'astroPH': load_subject_data('astroPH'), 'csLG': load_subject_data('csLG') } # Update existing functions to use the selected subject's data def get_embedding(text: Optional[str], model: str = EMBEDDING_MODEL) -> Optional[np.ndarray]: try: embedding = client.embeddings.create(input=[text], model=model).data[0].embedding return np.array(embedding, dtype=np.float32) except Exception as e: print(f"Error getting embedding: {e}") return None def intervened_hidden_to_intervened_embedding(topk_indices, topk_values, ae): with torch.no_grad(): return ae.decode_sparse(topk_indices, topk_values) # Function definitions for feature activation, co-occurrence, styling, etc. def get_feature_activations(subject, feature_index, m=5, min_length=100): abstract_texts = subject_data[subject]['abstract_texts'] abstract_embeddings = subject_data[subject]['abstract_embeddings'] topk_indices = subject_data[subject]['topk_indices'] topk_values = subject_data[subject]['topk_values'] doc_ids = abstract_texts['doc_ids'] abstracts = abstract_texts['abstracts'] feature_mask = topk_indices == feature_index activated_indices = np.where(feature_mask.any(axis=1))[0] activation_values = np.where(feature_mask, topk_values, 0).max(axis=1) sorted_activated_indices = activated_indices[np.argsort(-activation_values[activated_indices])] top_m_abstracts = [] top_m_indices = [] for i in sorted_activated_indices: if len(abstracts[i]) > min_length: top_m_abstracts.append((doc_ids[i], abstracts[i], activation_values[i])) top_m_indices.append(i) if len(top_m_abstracts) == m: break return top_m_abstracts def calculate_co_occurrences(subject, target_index, n_features=9216): topk_indices = subject_data[subject]['topk_indices'] mask = np.any(topk_indices == target_index, axis=1) co_occurring_indices = topk_indices[mask].flatten() co_occurrences = Counter(co_occurring_indices) del co_occurrences[target_index] result = np.zeros(n_features, dtype=int) result[list(co_occurrences.keys())] = list(co_occurrences.values()) return result def style_dataframe(df: pd.DataFrame, is_top: bool) -> pd.DataFrame: cosine_values = df['Cosine similarity'].astype(float) min_val = cosine_values.min() max_val = cosine_values.max() def color_similarity(val): val = float(val) # Normalize the value between 0 and 1 if is_top: normalized_val = (val - min_val) / (max_val - min_val) else: # For bottom correlated, reverse the normalization normalized_val = (max_val - val) / (max_val - min_val) # Adjust the color intensity to avoid zero intensity color_intensity = 0.2 + (normalized_val * 0.8) # This ensures the range is from 0.2 to 1.0 if is_top: color = f'background-color: rgba(0, 255, 0, {color_intensity:.2f})' else: color = f'background-color: rgba(255, 0, 0, {color_intensity:.2f})' return color return df.style.applymap(color_similarity, subset=['Cosine similarity']) def get_feature_from_index(subject, index): feature = next((f for f in subject_data[subject]['feature_analysis'] if f['index'] == index), None) return feature def visualize_feature(subject, index): feature = next((f for f in subject_data[subject]['feature_analysis'] if f['index'] == index), None) if feature is None: return "Invalid feature index", None, None, None, None, None, None output = f"# {feature['label']}\n\n" output += f"* Pearson correlation: {feature['pearson_correlation']:.4f}\n\n" output += f"* Density: {feature['density']:.4f}\n\n" # Top m abstracts top_m_abstracts = get_feature_activations(subject, index) # Create dataframe for top abstracts with clickable links df_data = [] for doc_id, abstract, activation_value in top_m_abstracts: title = abstract.split('\n\n')[0] title = title.replace('[', '').replace(']', '') title = title.replace("'", "") title = title.replace('"', '') url_id = doc_id.replace('_arXiv.txt', '') if 'astro-ph' in url_id: url_id = url_id.split('astro-ph')[1] url = f"https://arxiv.org/abs/astro-ph/{url_id}" else: if '.' in doc_id: url = f"https://arxiv.org/abs/{url_id}" else: url = f"https://arxiv.org/abs/hep-ph/{url_id}" linked_title = f"[{title}]({url})" df_data.append({"Title": linked_title, "Activation value": activation_value}) df_top_abstracts = pd.DataFrame(df_data) styled_top_abstracts = df_top_abstracts.style.format({ "Activation value": "{:.4f}" }) # Activation value distribution topk_indices = subject_data[subject]['topk_indices'] topk_values = subject_data[subject]['topk_values'] activation_values = np.where(topk_indices == index, topk_values, 0).max(axis=1) fig2 = px.histogram(x=activation_values, nbins=50) fig2.update_layout( #title=f'{feature["label"]}', xaxis_title='Activation value', yaxis_title=None, yaxis_type='log', height=220, ) # Correlated features decoder = subject_data[subject]['decoder'] feature_vector = decoder[:, index] decoder_without_feature = np.delete(decoder, index, axis=1) cosine_similarities = np.dot(feature_vector, decoder_without_feature) / (np.linalg.norm(decoder_without_feature, axis=0) * np.linalg.norm(feature_vector)) topk = 5 topk_indices_cosine = np.argsort(-cosine_similarities)[:topk] topk_values_cosine = cosine_similarities[topk_indices_cosine] bottomk = 5 bottomk_indices_cosine = np.argsort(cosine_similarities)[:bottomk] bottomk_values_cosine = cosine_similarities[bottomk_indices_cosine] df_top_correlated = pd.DataFrame({ "Feature": [get_feature_from_index(subject, i)['label'] for i in topk_indices_cosine], "Cosine similarity": topk_values_cosine }) df_top_correlated_styled = style_dataframe(df_top_correlated, is_top=True) # Create dataframe for bottom 5 correlated features df_bottom_correlated = pd.DataFrame({ "Feature": [get_feature_from_index(subject, i)['label'] for i in bottomk_indices_cosine], "Cosine similarity": bottomk_values_cosine }) df_bottom_correlated_styled = style_dataframe(df_bottom_correlated, is_top=False) # Co-occurrences co_occurrences = calculate_co_occurrences(subject, index) topk = 5 topk_indices_co_occurrence = np.argsort(-co_occurrences)[:topk] topk_values_co_occurrence = co_occurrences[topk_indices_co_occurrence] # Create dataframe for top 5 co-occurring features df_co_occurrences = pd.DataFrame({ "Feature": [get_feature_from_index(subject, i)['label'] for i in topk_indices_co_occurrence], "Co-occurrences": topk_values_co_occurrence }) df_co_occurrences_styled = df_co_occurrences.style.format({ "Co-occurrences": "{:.0f}" # Keep as integer }) return output, styled_top_abstracts, df_top_correlated_styled, df_bottom_correlated_styled, df_co_occurrences_styled, fig2 # Modify the main interface function def create_interface(): custom_css = """ #custom-slider-* { background-color: #ffe6e6; } """ with gr.Blocks(css=custom_css) as demo: subject = gr.Dropdown(choices=['astroPH', 'csLG'], label="Select Subject", value='astroPH') with gr.Tabs(): with gr.Tab("Home"): gr.Markdown(""" # SAErch: Sparse Autoencoder-enhanced Semantic Search Welcome to SAErch, an innovative approach to semantic search using Sparse Autoencoders (SAEs) trained on dense text embeddings. This tool builds upon recent advancements in the application of SAEs to language models and embeddings. ## Key Concepts: 1. **Sparse Autoencoders (SAEs)**: Neural networks that learn to reconstruct input data using a sparse set of features, helping to disentangle complex representations. SAEs have shown promising results in uncovering interpretable features in language models. 2. **Feature Families**: Groups of related SAE features that represent concepts at varying levels of abstraction, allowing for multi-scale semantic analysis and manipulation. 3. **Embedding Interventions**: Technique to modify search queries by manipulating specific semantic features identified by the SAE, enabling fine-grained control over query semantics. ## How It Works: 1. SAEs are trained on embeddings from scientific paper abstracts, learning interpretable features that capture various semantic concepts. 2. Users can interact with these features to fine-tune search queries. 3. The system performs semantic search using the modified embeddings, allowing for more precise and controllable results. ## Key References: - [Towards Monosemanticity: Decomposing Language Models With Dictionary Learning](https://transformer-circuits.pub/2023/monosemantic-features) - Anthropic's pioneering work on applying SAEs to language models. - [Prism: Mapping Interpretable Concepts and Features in a Latent Space of Language](https://thesephist.com/posts/prism/#caveats-and-limitations) - An early application of SAEs to embeddings, demonstrating their potential for interpretable concept mapping. - [Scaling and Evaluating Sparse Autoencoders](https://arxiv.org/html/2406.04093v1) - OpenAI's research on scaling SAEs, showcasing the effectiveness of top-k SAEs. Explore the "SAErch" tab to try out the semantic search capabilities, or dive into the "Feature Visualisation" tab to examine the learned features in more detail. This tool demonstrates how SAEs can bridge the gap between the semantic richness of dense embeddings and the interpretability of sparse representations, offering new possibilities for precise and explainable semantic search. """) with gr.Tab("SAErch"): input_text = gr.Textbox(label="input") search_results_state = gr.State([]) feature_values_state = gr.State([]) feature_indices_state = gr.State([]) manually_added_features_state = gr.State([]) def update_search_results(feature_values, feature_indices, manually_added_features, current_subject): ae = subject_data[current_subject]['ae'] abstract_embeddings = subject_data[current_subject]['abstract_embeddings'] abstract_texts = subject_data[current_subject]['abstract_texts'] df_metadata = subject_data[current_subject]['df_metadata'] # Combine manually added features with query-generated features all_indices = [] all_values = [] # Add manually added features first for index in manually_added_features: if index not in all_indices: all_indices.append(index) all_values.append(feature_values[feature_indices.index(index)] if index in feature_indices else 0.0) # Add remaining query-generated features for index, value in zip(feature_indices, feature_values): if index not in all_indices: all_indices.append(index) all_values.append(value) # Reconstruct query embedding topk_indices = torch.tensor(all_indices).to(device) topk_values = torch.tensor(all_values).to(device) intervened_embedding = intervened_hidden_to_intervened_embedding(topk_indices, topk_values, ae) intervened_embedding = intervened_embedding.cpu().numpy().flatten() # Perform similarity search sims = np.dot(abstract_embeddings, intervened_embedding) topk_indices_search = np.argsort(sims)[::-1][:10] doc_ids = abstract_texts['doc_ids'] topk_doc_ids = [doc_ids[i] for i in topk_indices_search] # Prepare search results search_results = [] for doc_id in topk_doc_ids: metadata = df_metadata[df_metadata['arxiv_id'] == doc_id].iloc[0] title = metadata['title'].replace('[', '').replace(']', '') title = title.replace("'", "") url_id = doc_id.replace('_arXiv.txt', '') if 'astro-ph' in url_id: url_id = url_id.split('astro-ph')[1] url = f"https://arxiv.org/abs/astro-ph/{url_id}" else: if '.' in doc_id: url = f"https://arxiv.org/abs/{doc_id.replace('_arXiv.txt', '')}" else: url = f"https://arxiv.org/abs/hep-ph/{doc_id.replace('_arXiv.txt', '')}" linked_title = f"[{title}]({url})" search_results.append([ linked_title, int(metadata['citation_count']), int(metadata['year']) ]) # Convert search_results to a DataFrame and apply styling df_search_results = pd.DataFrame(search_results, columns=["Title", "Citation Count", "Year"]) styled_search_results = df_search_results.style.format({ "Citation Count": "{:.0f}", # Keep as integer "Year": "{:.0f}" # Keep as integer }) return styled_search_results, all_values, all_indices @gr.render(inputs=[input_text, search_results_state, feature_values_state, feature_indices_state, manually_added_features_state, subject]) def show_components(text, search_results, feature_values, feature_indices, manually_added_features, current_subject): if len(text) == 0: return gr.Markdown("## No Input Provided") if not search_results or text != getattr(show_components, 'last_query', None): show_components.last_query = text query_embedding = get_embedding(text) ae = subject_data[current_subject]['ae'] with torch.no_grad(): recons, z_dict = ae(torch.tensor(query_embedding).unsqueeze(0).to(device)) topk_indices = z_dict['topk_indices'][0].cpu().numpy() topk_values = z_dict['topk_values'][0].cpu().numpy() feature_values = topk_values.tolist() feature_indices = topk_indices.tolist() search_results, feature_values, feature_indices = update_search_results(feature_values, feature_indices, manually_added_features, current_subject) with gr.Row(): with gr.Column(scale=2): df = gr.Dataframe( headers=["Title", "Citation Count", "Year"], value=search_results, label="Top 10 Search Results", datatype=["markdown", "number", "number"], # Add this line wrap=True # Add this line to ensure long titles don't get cut off ) feature_search = gr.Textbox(label="Search Feature Labels") feature_matches = gr.CheckboxGroup(label="Matching Features", choices=[]) add_button = gr.Button("Add Selected Features") def search_feature_labels(search_text): if not search_text: return gr.CheckboxGroup(choices=[]) matches = [f"{f['label']} ({f['index']})" for f in subject_data[current_subject]['feature_analysis'] if search_text.lower() in f['label'].lower()] return gr.CheckboxGroup(choices=matches[:10]) feature_search.change(search_feature_labels, inputs=[feature_search], outputs=[feature_matches]) def on_add_features(selected_features, current_values, current_indices, manually_added_features): if selected_features: new_indices = [int(f.split('(')[-1].strip(')')) for f in selected_features] # Add new indices to manually_added_features if they're not already there manually_added_features = list(dict.fromkeys(manually_added_features + new_indices)) return gr.CheckboxGroup(value=[]), current_values, current_indices, manually_added_features return gr.CheckboxGroup(value=[]), current_values, current_indices, manually_added_features add_button.click( on_add_features, inputs=[feature_matches, feature_values_state, feature_indices_state, manually_added_features_state], outputs=[feature_matches, feature_values_state, feature_indices_state, manually_added_features_state] ) with gr.Column(scale=1): update_button = gr.Button("Update Results") sliders = [] for i, (value, index) in enumerate(zip(feature_values, feature_indices)): feature = next((f for f in subject_data[current_subject]['feature_analysis'] if f['index'] == index), None) label = f"{feature['label']} ({index})" if feature else f"Feature {index}" # Transform the value to a 0-1 range transformed_value = max(0.01, min(1, value)) # Ensure value is between 0.01 and 1 linear_value = (np.log10(transformed_value) + 2) / 2 # Map 0.01-1 to 0-1 # Add prefix and change color for manually added features if index in manually_added_features: label = f"[Custom] {label}" slider = gr.Slider(minimum=0, maximum=1, step=0.01, value=linear_value, label=label, key=f"slider-{index}", elem_id=f"custom-slider-{index}") else: slider = gr.Slider(minimum=0, maximum=1, step=0.01, value=linear_value, label=label, key=f"slider-{index}") sliders.append(slider) def on_slider_change(*values): manually_added_features = values[-1] slider_values = list(values[:-1]) # Transform slider values back to original scale transformed_values = [10 ** ((2 * float(v)) - 2) for v in slider_values] # Reconstruct feature_indices based on the order of sliders reconstructed_indices = [int(slider.label.split('(')[-1].split(')')[0]) for slider in sliders] new_results, new_values, new_indices = update_search_results(transformed_values, reconstructed_indices, manually_added_features, current_subject) return new_results, new_values, new_indices, manually_added_features update_button.click( on_slider_change, inputs=sliders + [manually_added_features_state], outputs=[search_results_state, feature_values_state, feature_indices_state, manually_added_features_state] ) return [df, feature_search, feature_matches, add_button, update_button] + sliders with gr.Tab("Feature Visualisation"): gr.Markdown("# Feature Visualiser") with gr.Tabs(): with gr.Tab("Individual Features"): with gr.Row(): feature_search = gr.Textbox(label="Search Feature Labels") feature_matches = gr.CheckboxGroup(label="Matching Features", choices=[]) visualize_button = gr.Button("Visualize Feature") feature_info = gr.Markdown() abstracts_heading = gr.Markdown("## Top 5 Abstracts") top_abstracts = gr.Dataframe( headers=["Title", "Activation value"], datatype=["markdown", "number"], interactive=False, wrap=True ) gr.Markdown("## Correlated Features") with gr.Row(): with gr.Column(scale=1): gr.Markdown("### Top 5 Correlated Features") top_correlated = gr.Dataframe( headers=["Feature", "Cosine similarity"], interactive=False ) with gr.Column(scale=1): gr.Markdown("### Bottom 5 Correlated Features") bottom_correlated = gr.Dataframe( headers=["Feature", "Cosine similarity"], interactive=False ) with gr.Row(): with gr.Column(scale=1): gr.Markdown("## Top 5 Co-occurring Features") co_occurring_features = gr.Dataframe( headers=["Feature", "Co-occurrences"], interactive=False ) with gr.Column(scale=1): gr.Markdown(f"## Activation Value Distribution") activation_dist = gr.Plot() def search_feature_labels(search_text, current_subject): if not search_text: return gr.CheckboxGroup(choices=[]) matches = [f"{f['label']} ({f['index']})" for f in subject_data[current_subject]['feature_analysis'] if search_text.lower() in f['label'].lower()] return gr.CheckboxGroup(choices=matches[:10]) feature_search.change(search_feature_labels, inputs=[feature_search, subject], outputs=[feature_matches]) def on_visualize(selected_features, current_subject): if not selected_features: return "Please select a feature to visualize.", None, None, None, None, None, "", [] # Extract the feature index from the selected feature string feature_index = int(selected_features[0].split('(')[-1].strip(')')) feature_info, top_abstracts, top_correlated, bottom_correlated, co_occurring_features, activation_dist = visualize_feature(current_subject, feature_index) # Return the visualization results along with empty values for search box and checkbox return feature_info, top_abstracts, top_correlated, bottom_correlated, co_occurring_features, activation_dist, "", [] visualize_button.click( on_visualize, inputs=[feature_matches, subject], outputs=[feature_info, top_abstracts, top_correlated, bottom_correlated, co_occurring_features, activation_dist, feature_search, feature_matches] ) with gr.Tab("Feature Families"): gr.Markdown("# Feature Families") with gr.Row(): family_search = gr.Textbox(label="Search Feature Families") family_matches = gr.CheckboxGroup(label="Matching Feature Families", choices=[]) visualize_family_button = gr.Button("Visualize Feature Family") family_info = gr.Markdown() family_dataframe = gr.Dataframe( headers=["Feature", "F1 Score", "Pearson Correlation"], datatype=["markdown", "number", "number"], label="Family and Child Features" ) def search_feature_families(search_text, current_subject): family_analysis = subject_data[current_subject]['family_analysis'] if not search_text: return gr.CheckboxGroup(choices=[]) matches = [family['superfeature'] for family in family_analysis if search_text.lower() in family['superfeature'].lower()] return gr.CheckboxGroup(choices=matches[:10]) # Limit to top 10 matches def visualize_feature_family(selected_families, current_subject): if not selected_families: return "Please select a feature family to visualize.", None, "", [] selected_family = selected_families[0] # Take the first selected family family_analysis = subject_data[current_subject]['family_analysis'] family_data = next((family for family in family_analysis if family['superfeature'] == selected_family), None) if not family_data: return "Invalid feature family selected.", None, "", [] output = f"# {family_data['superfeature']}\n\n" # Create DataFrame df_data = [ { "Feature": f"## {family_data['superfeature']}", "F1 Score": round(family_data['family_f1'], 2), "Pearson Correlation": round(family_data['family_pearson'], 4) }, ] for name, f1, pearson in zip(family_data['feature_names'], family_data['feature_f1'], family_data['feature_pearson']): df_data.append({ "Feature": name, "F1 Score": round(f1, 2), "Pearson Correlation": round(pearson, 4) }) df = pd.DataFrame(df_data) # Add super reasoning below the dataframe output += "## Super Reasoning\n" output += f"{family_data['super_reasoning']}\n\n" return output, df, "", [] # Return empty string for search box and empty list for checkbox family_search.change(search_feature_families, inputs=[family_search, subject], outputs=[family_matches]) visualize_family_button.click( visualize_feature_family, inputs=[family_matches, subject], outputs=[family_info, family_dataframe, family_search, family_matches] ) # Add logic to update components when subject changes def on_subject_change(new_subject): # Clear all states and return empty values for all components return [], [], [], [], "", [], "", [], None, None, None, None, None, None subject.change( on_subject_change, inputs=[subject], outputs=[search_results_state, feature_values_state, feature_indices_state, manually_added_features_state, input_text, feature_matches, feature_search, feature_matches, feature_info, top_abstracts, top_correlated, bottom_correlated, co_occurring_features, activation_dist] ) return demo # Launch the interface if __name__ == "__main__": demo = create_interface() demo.launch()