├── src ├── plot_utils.py └── text_clustering.py ├── examples └── README.md ├── README.md ├── run_pipeline.py └── LICENSE /src/plot_utils.py: -------------------------------------------------------------------------------- 1 | import os 2 | 3 | import matplotlib.pyplot as plt 4 | import pandas as pd 5 | import seaborn as sns 6 | from datasets import load_dataset 7 | 8 | 9 | def get_size(ds): 10 | return len([e for i in range(len(ds)) for e in ds[i]["examples"]]) 11 | 12 | 13 | def extract_score(example): 14 | summary = example["summary"] 15 | category = summary.split(". Educational")[0].strip() 16 | score = summary.split(" Educational score: ")[1].strip() 17 | return {"category": category, "educational_score": score} 18 | 19 | 20 | def plot_distributions(ds_path, image_path="."): 21 | """Plot distribution of educational score of topics & distribution of samples accross topics""" 22 | ds = load_dataset(ds_path, split="train", num_proc=2, token=os.getenv("HF_TOKEN")) 23 | ds = ds.map(extract_score) 24 | print(ds["category"]) 25 | ds = ds.filter(lambda x: x["educational_score"] not in ["None", ""]) 26 | # distribution of scores 27 | df = ds.to_pandas() 28 | df["educational_score"] = pd.to_numeric(df["educational_score"], errors="coerce") 29 | df.dropna(subset=["educational_score"], inplace=True) 30 | 31 | sns.set_theme(style="whitegrid") 32 | plt.figure(figsize=(10, 6)) 33 | sns.histplot(df["educational_score"], kde=False, bins=10) 34 | plt.title("Distribution of Educational Scores") 35 | plt.xlabel("Educational Score") 36 | plt.ylabel("Frequency") 37 | plt.savefig(f"{image_path}/educational_score.png", bbox_inches="tight") 38 | 39 | # distribution of samples 40 | df = ds.to_pandas().explode("examples") 41 | sorted_filtered_ds = df.groupby(by="category").size().sort_values(ascending=False) 42 | category_df = sorted_filtered_ds.reset_index() 43 | category_df.columns = ["category", "number_files"] 44 | print(f"Saving csv in {image_path}!") 45 | category_df.to_csv(f"{image_path}/df_categories_count.csv") 46 | 47 | sns.set_theme(style="whitegrid") 48 | plt.figure(figsize=(25, 20)) 49 | 50 | barplot = sns.barplot( 51 | x="number_files", y="category", data=category_df, palette="Blues_d", ci=None 52 | ) 53 | 54 | plt.xlabel("Number of Examples") 55 | plt.ylabel("Categories") 56 | plt.title("Histogram of Categories and their number of FW files") 57 | plt.tight_layout(pad=1.0) 58 | plt.show() 59 | plt.savefig(f"{image_path}/topics_distpng", bbox_inches="tight", dpi=200) 60 | -------------------------------------------------------------------------------- /examples/README.md: -------------------------------------------------------------------------------- 1 | # Examples 2 | 3 | ## Cosmopedia experiments: clustering of web samples 4 | 5 | Here you can find the commands we used during the selection of web samples for [Cosmopedia](https://huggingface.co/datasets/HuggingFaceTB/cosmopedia) prompts. 6 | 7 | Our goal was to find the topics in random web samples and their educational score. The topics were used in the creation of prompts for synthetic data generation and helped us understand the range of domains covered. Initially, we clustered **100,000 samples**, yielding **145 clusters**. Then we assigned **15 million samples** to these clusters using the inference mode of `text-clustering`; however, half of them did not fit into any cluster and were excluded from prompt creation. 8 | 9 | For illustration, we will use [AutoMathText](https://huggingface.co/datasets/math-ai/AutoMathText) here. In Cosmopedia we used samples from a web dataset like [RefineWeb](https://huggingface.co/datasets/tiiuae/falcon-refinedweb). 10 | 11 | We will run the clustering using `topic_mode` single with educational scores. This pipeline clusters files and prompts an LLM (by default [Mixtral-8x7B-Instruct-v0.1](https://huggingface.co/mistralai/Mixtral-8x7B-Instruct-v0.1)) to find the topic of each cluster and give it an educational score. We plot the distribution of samples over topics and the distribution of the educational score and save the plots in the `save_load_path` folder. 12 | 13 | ```bash 14 | python run_pipeline.py --mode run \ 15 | --save_load_path './web_samples_100k' \ 16 | --input_dataset math-ai/AutoMathText \ 17 | --data_subset "web-0.70-to-1.00" \ 18 | --input_content text \ 19 | --n_samples 100000 \ 20 | --build_hf_ds \ 21 | --topic_mode single_topic \ 22 | --dbscan_eps 0.08 \ 23 | --dbscan_min_samples 50 24 | ``` 25 | 26 | 27 | This detects 213 clusters that you can visualize in this [plot](https://huggingface.co/datasets/HuggingFaceTB/miscellaneous/blob/main/AMT_plots/topics_distpng.png) along with the [educational scores](https://huggingface.co/datasets/HuggingFaceTB/miscellaneous/blob/main/AMT_plots/educational_score.png) which is very high for this AutoMathText dataset. 28 | 29 | When using general web datasets, you might want to filter out files with a lower quality by discarding clusters with a low educational score (e.g. Explicit Adult Content). You can check this [demo](https://huggingface.co/spaces/HuggingFaceTB/inspect_clusters_free_topics) for an example. 30 | 31 | 32 |
33 | clusters 34 |

The clusters of AutoMathText

35 |
36 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # Text Clustering 2 | 3 | The Text Clustering repository contains tools to easily embed and cluster texts as well as label clusters semantically. This repository is a work in progress and serves as a minimal codebase that can be modified and adapted to other use cases. 4 | 5 |
6 |
Clustering of texts in the Cosmopedia dataset.
7 | 8 | 9 | ## How it works 10 | The pipeline consists of several distinct blocks that can be customized and the whole pipeline can run in a few minutes on a consumer laptop. Each block uses existing standard methods and works quite robustly. 11 | 12 |
13 |
Text clustering pipeline.
14 | 15 | 16 | ## Install 17 | Install the following libraries to get started: 18 | ```bash 19 | pip install scikit-learn umap-learn sentence_transformers faiss-cpu plotly matplotlib datasets 20 | ``` 21 | Clone this repository and navigate to the folder: 22 | ```bash 23 | git clone https://github.com/huggingface/text-clustering.git 24 | cd text-clustering 25 | ``` 26 | 27 | ## Usage 28 | 29 | Run pipeline and visualize results: 30 | 31 | ```python 32 | from src.text_clustering import ClusterClassifier 33 | from datasets import load_dataset 34 | 35 | SAMPLE = 100_000 36 | 37 | texts = load_dataset("HuggingFaceTB/cosmopedia-100k", split="train").select(range(SAMPLE))["text"] 38 | 39 | cc = ClusterClassifier(embed_device="mps") 40 | 41 | # run the pipeline: 42 | embs, labels, summaries = cc.fit(texts) 43 | 44 | # show the results 45 | cc.show() 46 | 47 | # save 48 | cc.save("./cc_100k") 49 | ``` 50 | 51 | Load classifier and run inference: 52 | ```python 53 | from src.text_clustering import ClusterClassifier 54 | 55 | cc = ClusterClassifier(embed_device="mps") 56 | 57 | # load state 58 | cc.load("./cc_100k") 59 | 60 | # visualize 61 | cc.show() 62 | 63 | # classify new texts with k-nearest neighbour search 64 | cluster_labels, embeddings = cc.infer(some_texts, top_k=1) 65 | ``` 66 | 67 | If you want to reproduce the color scheme in the plot above you can add the following code before you run `cc.show()`: 68 | ```python 69 | from cycler import cycler 70 | import matplotlib.pyplot as plt 71 | 72 | default_cycler = (cycler(color=[ 73 | "0F0A0A", 74 | "FF6600", 75 | "FFBE00", 76 | "496767", 77 | "87A19E", 78 | "FF9200", 79 | "0F3538", 80 | "F8E08E", 81 | "0F2021", 82 | "FAFAF0"]) 83 | ) 84 | plt.rc('axes', prop_cycle=default_cycler) 85 | ``` 86 | If you would like to customize the plotting further the easiest way is to customize or overwrite the `_show_mpl` and `_show_plotly` methods. 87 | 88 | You can also run the pipeline using a script with: 89 | ```bash 90 | # run a new pipeline 91 | python run_pipeline.py --mode run --save_load_path './cc_100k' --n_samples 100000 --build_hf_ds 92 | # load existing pipeline 93 | python run_pipeline.py --mode load --save_load_path './cc_100k' --build_hf_ds 94 | # inference mode on new texts from an input dataset 95 | python run_pipeline.py --mode infer --save_load_path './cc_100k' --n_samples --input_dataset 96 | ``` 97 | The `build_hf_ds` flag builds and pushes HF datasets, for the files and clusters, that can be directly used in the FW visualization space. In `infer` mode, we push the clusters dataset by default. 98 | 99 | You can also change how the clusters are labeled (multiple topics (default) vs single topic with an educational score) using the flag `--topic_mode`. 100 | 101 | ## Examples 102 | 103 | Check the `examples` folder for an example of clustering and topic labeling applied to the [AutoMathText](https://huggingface.co/datasets/math-ai/AutoMathText/) dataset, utilizing [Cosmopedia](https://huggingface.co/datasets/HuggingFaceTB/cosmopedia)'s web labeling approach. -------------------------------------------------------------------------------- /run_pipeline.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | import textwrap 3 | import time 4 | 5 | import numpy as np 6 | import pandas as pd 7 | from datasets import Dataset, load_dataset 8 | 9 | from src.plot_utils import plot_distributions 10 | from src.text_clustering import ClusterClassifier 11 | 12 | INSTRUCTION_SINGLE_TOPIC = "The examples below are web samples from the same cluster, identify the topic they have in common, for example: Philosophy, Lifesyle, Linear Algebra, Biochemistry, Economics...\ 13 | Additionally determine if the topics in the examples \ 14 | are broadly suitable as college/school material, while being mindful to exclude any sensitive/inappropriate/irrelevant content, \ 15 | including but not limited to sex, explicit violence, ads & scams, and other non-academic subjects. Consider a wide range of content including scientific, \ 16 | educational, historical, cultural, and practical applications and give a rating of how educational these topics could be from 1 to 10, 1 being extremely un-educational \ 17 | and inapproriate for an education setting and 10 being highly educational. The output format should be like this: Topic: the_topic, Educational value rating: score." 18 | INSTRUCTION_MULTIPLE_TOPICS = "Use three words total (comma separated)\ 19 | to describe general topics in above texts. Under no circumstances use enumeration. \ 20 | Example format: Tree, Cat, Fireman" 21 | 22 | 23 | TEMPLATE_MULTIPLE_TOPICS = "[INST]{examples}\n\n{instruction}[/INST]" 24 | TEMPLATE_SINGLE_TOPIC = "[INST]{instruction}\n\nExamples:\n{examples}\nRemember that the output format should be like this: Topic: the_topic, Educational value rating: score.[/INST]" 25 | 26 | 27 | def get_args(): 28 | parser = argparse.ArgumentParser() 29 | parser.add_argument("--n_samples", type=int, default=100_000) 30 | parser.add_argument("--start", type=int, default=-1) 31 | parser.add_argument("--end", type=int, default=100_000) 32 | parser.add_argument("--device", type=str, default="cuda") 33 | parser.add_argument("--save_load_path", type=str, default="./cc_100k") 34 | parser.add_argument( 35 | "--input_dataset", 36 | type=str, 37 | default="HuggingFaceFW/FW-12-12-2023-CC-2023-06", 38 | help="dataset with the samples to use for clustering", 39 | ) 40 | parser.add_argument( 41 | "--data_subset", 42 | type=str, 43 | default=None, 44 | help="dataset subset", 45 | ) 46 | parser.add_argument("--input_content", type=str, default="content") 47 | parser.add_argument( 48 | "--topic_mode", 49 | type=str, 50 | choices=["single_topic", "multiple_topics"], 51 | default="multiple_topics", 52 | help="Specify 'single_topic' to generate only one topic and score its educational value, or 'multiple_topics' to generate the 3 most relevant topics in the cluster.", 53 | ) 54 | parser.add_argument( 55 | "--dbscan_eps", 56 | type=float, 57 | default=0.08, 58 | help="The maximum distance between two samples for them to be considered as in the neighborhood of each other.", 59 | ) 60 | parser.add_argument( 61 | "--dbscan_min_samples", 62 | type=int, 63 | default=50, 64 | help="The number of samples in a neighborhood for a point to be considered as a core point.", 65 | ) 66 | parser.add_argument( 67 | "--mode", 68 | choices=["run", "load", "infer"], 69 | default="run", 70 | help="Run the pipeline from scratch/load existing model to build hf datasets or to infer on new texts", 71 | ) 72 | parser.add_argument( 73 | "--inference_repo_name", 74 | type=str, 75 | default="infer_fw_on_ultrachat", 76 | help="HF repo name for the clusters dataset in inference mode", 77 | ) 78 | parser.add_argument( 79 | "--build_hf_ds", 80 | action="store_true", 81 | help="Builds HF datasets used for space visualization and pushes them to the hub", 82 | ) 83 | parser.add_argument("--username", type=str, default="loubnabnl") 84 | return parser.parse_args() 85 | 86 | 87 | def extract_res(example): 88 | summary = example["summary"] 89 | category = summary.split(". Educational")[0].strip() 90 | score = summary.split(" Educational score: ")[1].strip() 91 | return {"category": category, "educational_score": score} 92 | 93 | 94 | def build_hf_data_clusters(cc, texts=None, labels=None): 95 | """ 96 | Build an HF dataset containing information on each cluster. 97 | 98 | Args: 99 | cc: ClusterClassifier object. 100 | texts: list of texts used for inference mode. 101 | labels: list of cluster labels corresponding to the texts for inference mode. 102 | 103 | If `texts` and `labels` are not provided, the function will use the data available in `cc` 104 | to construct the dataset. Otherwise it will run in inference mode on texts. 105 | """ 106 | cluster_data = [] 107 | for cluster_id in cc.label2docs.keys(): 108 | if cluster_id == -1: 109 | continue 110 | 111 | # inference mode 112 | if texts is not None and labels is not None: 113 | labels_array = np.array(labels) 114 | files_in_cluster = np.where(labels_array == cluster_id)[0] 115 | examples = [texts[doc_id] for doc_id in files_in_cluster] 116 | else: 117 | doc_ids = cc.label2docs[cluster_id] 118 | examples = [cc.texts[doc_id] for doc_id in doc_ids] 119 | 120 | cluster_info = { 121 | "cluster_id": cluster_id, 122 | "summary": cc.cluster_summaries[cluster_id], 123 | "examples": examples, 124 | } 125 | 126 | if not texts: 127 | cluster_info["position"] = cc.cluster_centers[cluster_id] 128 | 129 | cluster_data.append(cluster_info) 130 | 131 | return Dataset.from_pandas(pd.DataFrame(cluster_data)) 132 | 133 | 134 | def build_hf_data_files(cc): 135 | """ 136 | Build an HF dataset containing information on each file and the cluster they belong to 137 | """ 138 | 139 | df = pd.DataFrame( 140 | data={ 141 | "X": cc.projections[:, 0], 142 | "Y": cc.projections[:, 1], 143 | "labels": cc.cluster_labels, 144 | "content_display": [textwrap.fill(txt[:1024], 64) for txt in cc.texts], 145 | } 146 | ) 147 | return Dataset.from_pandas(df) 148 | 149 | 150 | def build_and_push(cc, args): 151 | """Build HF files & clusters datasts and push them to the hub""" 152 | print("Building HF datasets...") 153 | ds = build_hf_data_clusters(cc) 154 | ds = ds.map(extract_res) 155 | data_clusters = build_hf_data_files(cc) 156 | print(f"Files dataset {ds}\nClusters dataset {data_clusters}") 157 | 158 | repo_name = args.save_load_path.split("/")[-1] 159 | print(f"Pushing to the hub at {repo_name}...") 160 | ds.push_to_hub(f"{args.username}/{repo_name}", private=True) 161 | data_clusters.push_to_hub(f"{args.username}/{repo_name}_clusters", private=True) 162 | 163 | 164 | def main(): 165 | args = get_args() 166 | 167 | template = ( 168 | TEMPLATE_MULTIPLE_TOPICS 169 | if args.topic_mode == "multiple_topics" 170 | else TEMPLATE_SINGLE_TOPIC 171 | ) 172 | instruction = ( 173 | INSTRUCTION_MULTIPLE_TOPICS 174 | if args.topic_mode == "multiple_topics" 175 | else INSTRUCTION_SINGLE_TOPIC 176 | ) 177 | print(f"Using {args.topic_mode} for topic labeling") 178 | cc = ClusterClassifier( 179 | embed_device=args.device, 180 | topic_mode=args.topic_mode, 181 | summary_template=template, 182 | summary_instruction=instruction, 183 | dbscan_eps=args.dbscan_eps, 184 | dbscan_min_samples=args.dbscan_min_samples, 185 | ) 186 | 187 | if args.mode == "run": 188 | # Run a new pipeline on texts 189 | dataset_args = (args.input_dataset, args.data_subset) if args.data_subset else (args.input_dataset,) 190 | ds = load_dataset(*dataset_args, split="train", token=True).shuffle( 191 | seed=42 192 | ) 193 | 194 | print(ds) 195 | indexes = ( 196 | range(args.start, args.end) if args.start > 0 else range(args.n_samples) 197 | ) 198 | text_start = f" starting from {args.start}" if args.start > 0 else "" 199 | print(f"Processing {len(indexes)} samples{text_start}") 200 | 201 | texts = ds.select(indexes)[args.input_content] 202 | 203 | _, _, summaries = cc.fit(texts) 204 | print(f"10 example Summaries:\n{[e for e in summaries.values()][:10]}") 205 | 206 | cc.save(args.save_load_path) 207 | print(f"Saved clusters in {args.save_load_path}.") 208 | 209 | if args.build_hf_ds: 210 | build_and_push(cc, args) 211 | 212 | ds_path = f"{args.username}/{args.save_load_path.split('/')[-1]}" 213 | if args.topic_mode == "single_topic": 214 | plot_distributions(ds_path, image_path=args.save_load_path) 215 | print("📊 Saved plots for educational score and files distribution.") 216 | 217 | elif args.mode == "infer": 218 | # Run inference mode on texts using an existing pipeline 219 | cc.load(args.save_load_path) 220 | indexes = ( 221 | range(args.start, args.end) if args.start >= 0 else range(args.n_samples) 222 | ) 223 | text_start = f" starting from {args.start}" if args.start >= 0 else "" 224 | print( 225 | f"Running inference on {len(indexes)} samples{text_start} of {args.input_dataset} using clusters in {args.save_load_path}." 226 | ) 227 | dataset_args = (args.input_dataset, args.data_subset) if args.data_subset else (args.input_dataset,) 228 | ds = load_dataset(*dataset_args, split="train", token=True) 229 | texts = ds.select(indexes)[args.input_content] 230 | 231 | start_time = time.time() 232 | cluster_labels, _ = cc.infer(texts, top_k=1) 233 | 234 | ds = build_hf_data_clusters(cc, texts, cluster_labels) 235 | print(f"Total time is {(time.time() - start_time)/60}min") 236 | target_repo = f"{args.username}/{args.inference_repo_name}" 237 | print(f"Samples with clusters: {ds}") 238 | print(f"Pushing to hub at {target_repo}...") 239 | ds.push_to_hub(f"{target_repo}", private=True) 240 | 241 | else: 242 | # Load existing pipeline 243 | if args.build_hf_ds: 244 | cc.load(args.save_load_path) 245 | build_and_push(cc, args) 246 | ds_path = f"{args.username}/{args.save_load_path.split('/')[-1]}" 247 | if args.topic_mode == "single_topic": 248 | plot_distributions(ds_path, image_path=args.save_load_path) 249 | print("📊 Saved plots for educational score and files distribution.") 250 | else: 251 | print("Using mode=load but build_hf_ds is False, nothing to be done.") 252 | 253 | print("Done 🎉") 254 | 255 | 256 | if __name__ == "__main__": 257 | main() 258 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | Apache License 2 | Version 2.0, January 2004 3 | http://www.apache.org/licenses/ 4 | 5 | TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION 6 | 7 | 1. Definitions. 8 | 9 | "License" shall mean the terms and conditions for use, reproduction, 10 | and distribution as defined by Sections 1 through 9 of this document. 11 | 12 | "Licensor" shall mean the copyright owner or entity authorized by 13 | the copyright owner that is granting the License. 14 | 15 | "Legal Entity" shall mean the union of the acting entity and all 16 | other entities that control, are controlled by, or are under common 17 | control with that entity. For the purposes of this definition, 18 | "control" means (i) the power, direct or indirect, to cause the 19 | direction or management of such entity, whether by contract or 20 | otherwise, or (ii) ownership of fifty percent (50%) or more of the 21 | outstanding shares, or (iii) beneficial ownership of such entity. 22 | 23 | "You" (or "Your") shall mean an individual or Legal Entity 24 | exercising permissions granted by this License. 25 | 26 | "Source" form shall mean the preferred form for making modifications, 27 | including but not limited to software source code, documentation 28 | source, and configuration files. 29 | 30 | "Object" form shall mean any form resulting from mechanical 31 | transformation or translation of a Source form, including but 32 | not limited to compiled object code, generated documentation, 33 | and conversions to other media types. 34 | 35 | "Work" shall mean the work of authorship, whether in Source or 36 | Object form, made available under the License, as indicated by a 37 | copyright notice that is included in or attached to the work 38 | (an example is provided in the Appendix below). 39 | 40 | "Derivative Works" shall mean any work, whether in Source or Object 41 | form, that is based on (or derived from) the Work and for which the 42 | editorial revisions, annotations, elaborations, or other modifications 43 | represent, as a whole, an original work of authorship. For the purposes 44 | of this License, Derivative Works shall not include works that remain 45 | separable from, or merely link (or bind by name) to the interfaces of, 46 | the Work and Derivative Works thereof. 47 | 48 | "Contribution" shall mean any work of authorship, including 49 | the original version of the Work and any modifications or additions 50 | to that Work or Derivative Works thereof, that is intentionally 51 | submitted to Licensor for inclusion in the Work by the copyright owner 52 | or by an individual or Legal Entity authorized to submit on behalf of 53 | the copyright owner. For the purposes of this definition, "submitted" 54 | means any form of electronic, verbal, or written communication sent 55 | to the Licensor or its representatives, including but not limited to 56 | communication on electronic mailing lists, source code control systems, 57 | and issue tracking systems that are managed by, or on behalf of, the 58 | Licensor for the purpose of discussing and improving the Work, but 59 | excluding communication that is conspicuously marked or otherwise 60 | designated in writing by the copyright owner as "Not a Contribution." 61 | 62 | "Contributor" shall mean Licensor and any individual or Legal Entity 63 | on behalf of whom a Contribution has been received by Licensor and 64 | subsequently incorporated within the Work. 65 | 66 | 2. Grant of Copyright License. Subject to the terms and conditions of 67 | this License, each Contributor hereby grants to You a perpetual, 68 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 69 | copyright license to reproduce, prepare Derivative Works of, 70 | publicly display, publicly perform, sublicense, and distribute the 71 | Work and such Derivative Works in Source or Object form. 72 | 73 | 3. Grant of Patent License. Subject to the terms and conditions of 74 | this License, each Contributor hereby grants to You a perpetual, 75 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 76 | (except as stated in this section) patent license to make, have made, 77 | use, offer to sell, sell, import, and otherwise transfer the Work, 78 | where such license applies only to those patent claims licensable 79 | by such Contributor that are necessarily infringed by their 80 | Contribution(s) alone or by combination of their Contribution(s) 81 | with the Work to which such Contribution(s) was submitted. If You 82 | institute patent litigation against any entity (including a 83 | cross-claim or counterclaim in a lawsuit) alleging that the Work 84 | or a Contribution incorporated within the Work constitutes direct 85 | or contributory patent infringement, then any patent licenses 86 | granted to You under this License for that Work shall terminate 87 | as of the date such litigation is filed. 88 | 89 | 4. Redistribution. You may reproduce and distribute copies of the 90 | Work or Derivative Works thereof in any medium, with or without 91 | modifications, and in Source or Object form, provided that You 92 | meet the following conditions: 93 | 94 | (a) You must give any other recipients of the Work or 95 | Derivative Works a copy of this License; and 96 | 97 | (b) You must cause any modified files to carry prominent notices 98 | stating that You changed the files; and 99 | 100 | (c) You must retain, in the Source form of any Derivative Works 101 | that You distribute, all copyright, patent, trademark, and 102 | attribution notices from the Source form of the Work, 103 | excluding those notices that do not pertain to any part of 104 | the Derivative Works; and 105 | 106 | (d) If the Work includes a "NOTICE" text file as part of its 107 | distribution, then any Derivative Works that You distribute must 108 | include a readable copy of the attribution notices contained 109 | within such NOTICE file, excluding those notices that do not 110 | pertain to any part of the Derivative Works, in at least one 111 | of the following places: within a NOTICE text file distributed 112 | as part of the Derivative Works; within the Source form or 113 | documentation, if provided along with the Derivative Works; or, 114 | within a display generated by the Derivative Works, if and 115 | wherever such third-party notices normally appear. The contents 116 | of the NOTICE file are for informational purposes only and 117 | do not modify the License. You may add Your own attribution 118 | notices within Derivative Works that You distribute, alongside 119 | or as an addendum to the NOTICE text from the Work, provided 120 | that such additional attribution notices cannot be construed 121 | as modifying the License. 122 | 123 | You may add Your own copyright statement to Your modifications and 124 | may provide additional or different license terms and conditions 125 | for use, reproduction, or distribution of Your modifications, or 126 | for any such Derivative Works as a whole, provided Your use, 127 | reproduction, and distribution of the Work otherwise complies with 128 | the conditions stated in this License. 129 | 130 | 5. Submission of Contributions. Unless You explicitly state otherwise, 131 | any Contribution intentionally submitted for inclusion in the Work 132 | by You to the Licensor shall be under the terms and conditions of 133 | this License, without any additional terms or conditions. 134 | Notwithstanding the above, nothing herein shall supersede or modify 135 | the terms of any separate license agreement you may have executed 136 | with Licensor regarding such Contributions. 137 | 138 | 6. Trademarks. This License does not grant permission to use the trade 139 | names, trademarks, service marks, or product names of the Licensor, 140 | except as required for reasonable and customary use in describing the 141 | origin of the Work and reproducing the content of the NOTICE file. 142 | 143 | 7. Disclaimer of Warranty. Unless required by applicable law or 144 | agreed to in writing, Licensor provides the Work (and each 145 | Contributor provides its Contributions) on an "AS IS" BASIS, 146 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or 147 | implied, including, without limitation, any warranties or conditions 148 | of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A 149 | PARTICULAR PURPOSE. You are solely responsible for determining the 150 | appropriateness of using or redistributing the Work and assume any 151 | risks associated with Your exercise of permissions under this License. 152 | 153 | 8. Limitation of Liability. In no event and under no legal theory, 154 | whether in tort (including negligence), contract, or otherwise, 155 | unless required by applicable law (such as deliberate and grossly 156 | negligent acts) or agreed to in writing, shall any Contributor be 157 | liable to You for damages, including any direct, indirect, special, 158 | incidental, or consequential damages of any character arising as a 159 | result of this License or out of the use or inability to use the 160 | Work (including but not limited to damages for loss of goodwill, 161 | work stoppage, computer failure or malfunction, or any and all 162 | other commercial damages or losses), even if such Contributor 163 | has been advised of the possibility of such damages. 164 | 165 | 9. Accepting Warranty or Additional Liability. While redistributing 166 | the Work or Derivative Works thereof, You may choose to offer, 167 | and charge a fee for, acceptance of support, warranty, indemnity, 168 | or other liability obligations and/or rights consistent with this 169 | License. However, in accepting such obligations, You may act only 170 | on Your own behalf and on Your sole responsibility, not on behalf 171 | of any other Contributor, and only if You agree to indemnify, 172 | defend, and hold each Contributor harmless for any liability 173 | incurred by, or claims asserted against, such Contributor by reason 174 | of your accepting any such warranty or additional liability. 175 | 176 | END OF TERMS AND CONDITIONS 177 | 178 | APPENDIX: How to apply the Apache License to your work. 179 | 180 | To apply the Apache License to your work, attach the following 181 | boilerplate notice, with the fields enclosed by brackets "[]" 182 | replaced with your own identifying information. (Don't include 183 | the brackets!) The text should be enclosed in the appropriate 184 | comment syntax for the file format. We also recommend that a 185 | file or class name and description of purpose be included on the 186 | same "printed page" as the copyright notice for easier 187 | identification within third-party archives. 188 | 189 | Copyright 2023 Hugging Face 190 | 191 | Licensed under the Apache License, Version 2.0 (the "License"); 192 | you may not use this file except in compliance with the License. 193 | You may obtain a copy of the License at 194 | 195 | http://www.apache.org/licenses/LICENSE-2.0 196 | 197 | Unless required by applicable law or agreed to in writing, software 198 | distributed under the License is distributed on an "AS IS" BASIS, 199 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 200 | See the License for the specific language governing permissions and 201 | limitations under the License. 202 | -------------------------------------------------------------------------------- /src/text_clustering.py: -------------------------------------------------------------------------------- 1 | import json 2 | import logging 3 | import os 4 | import random 5 | import textwrap 6 | from collections import Counter, defaultdict 7 | 8 | import faiss 9 | import matplotlib.pyplot as plt 10 | import numpy as np 11 | import pandas as pd 12 | import plotly.express as px 13 | from huggingface_hub import InferenceClient 14 | from sentence_transformers import SentenceTransformer 15 | from sklearn.cluster import DBSCAN 16 | from tqdm import tqdm 17 | from umap import UMAP 18 | 19 | logging.basicConfig(level=logging.INFO) 20 | 21 | 22 | DEFAULT_INSTRUCTION = ( 23 | instruction 24 | ) = "Use three words total (comma separated)\ 25 | to describe general topics in above texts. Under no circumstances use enumeration. \ 26 | Example format: Tree, Cat, Fireman" 27 | 28 | DEFAULT_TEMPLATE = "[INST]{examples}\n\n{instruction}[/INST]" 29 | 30 | 31 | class ClusterClassifier: 32 | def __init__( 33 | self, 34 | embed_model_name="all-MiniLM-L6-v2", 35 | embed_device="cpu", 36 | embed_batch_size=64, 37 | embed_max_seq_length=512, 38 | embed_agg_strategy=None, 39 | umap_components=2, 40 | umap_metric="cosine", 41 | dbscan_eps=0.08, 42 | dbscan_min_samples=50, 43 | dbscan_n_jobs=16, 44 | summary_create=True, 45 | summary_model="mistralai/Mixtral-8x7B-Instruct-v0.1", 46 | topic_mode="multiple_topics", 47 | summary_n_examples=10, 48 | summary_chunk_size=420, 49 | summary_model_token=True, 50 | summary_template=None, 51 | summary_instruction=None, 52 | ): 53 | self.embed_model_name = embed_model_name 54 | self.embed_device = embed_device 55 | self.embed_batch_size = embed_batch_size 56 | self.embed_max_seq_length = embed_max_seq_length 57 | self.embed_agg_strategy = embed_agg_strategy 58 | 59 | self.umap_components = umap_components 60 | self.umap_metric = umap_metric 61 | 62 | self.dbscan_eps = dbscan_eps 63 | self.dbscan_min_samples = dbscan_min_samples 64 | self.dbscan_n_jobs = dbscan_n_jobs 65 | 66 | self.summary_create = summary_create 67 | self.summary_model = summary_model 68 | self.topic_mode = topic_mode 69 | self.summary_n_examples = summary_n_examples 70 | self.summary_chunk_size = summary_chunk_size 71 | self.summary_model_token = summary_model_token 72 | 73 | if summary_template is None: 74 | self.summary_template = DEFAULT_TEMPLATE 75 | else: 76 | self.summary_template = summary_template 77 | 78 | if summary_instruction is None: 79 | self.summary_instruction = DEFAULT_INSTRUCTION 80 | else: 81 | self.summary_instruction = summary_instruction 82 | 83 | self.embeddings = None 84 | self.faiss_index = None 85 | self.cluster_labels = None 86 | self.texts = None 87 | self.projections = None 88 | self.umap_mapper = None 89 | self.id2label = None 90 | self.label2docs = None 91 | 92 | self.embed_model = SentenceTransformer( 93 | self.embed_model_name, device=self.embed_device 94 | ) 95 | self.embed_model.max_seq_length = self.embed_max_seq_length 96 | 97 | def fit(self, texts, embeddings=None): 98 | self.texts = texts 99 | 100 | if embeddings is None: 101 | logging.info("embedding texts...") 102 | self.embeddings = self.embed(texts) 103 | else: 104 | logging.info("using precomputed embeddings...") 105 | self.embeddings = embeddings 106 | 107 | logging.info("building faiss index...") 108 | self.faiss_index = self.build_faiss_index(self.embeddings) 109 | logging.info("projecting with umap...") 110 | self.projections, self.umap_mapper = self.project(self.embeddings) 111 | logging.info("dbscan clustering...") 112 | self.cluster_labels = self.cluster(self.projections) 113 | 114 | self.id2cluster = { 115 | index: label for index, label in enumerate(self.cluster_labels) 116 | } 117 | self.label2docs = defaultdict(list) 118 | for i, label in enumerate(self.cluster_labels): 119 | self.label2docs[label].append(i) 120 | 121 | self.cluster_centers = {} 122 | for label in self.label2docs.keys(): 123 | x = np.mean([self.projections[doc, 0] for doc in self.label2docs[label]]) 124 | y = np.mean([self.projections[doc, 1] for doc in self.label2docs[label]]) 125 | self.cluster_centers[label] = (x, y) 126 | 127 | if self.summary_create: 128 | logging.info("summarizing cluster centers...") 129 | self.cluster_summaries = self.summarize(self.texts, self.cluster_labels) 130 | else: 131 | self.cluster_summaries = None 132 | 133 | return self.embeddings, self.cluster_labels, self.cluster_summaries 134 | 135 | def infer(self, texts, top_k=1): 136 | embeddings = self.embed(texts) 137 | 138 | dist, neighbours = self.faiss_index.search(embeddings, top_k) 139 | inferred_labels = [] 140 | for i in tqdm(range(embeddings.shape[0])): 141 | labels = [self.cluster_labels[doc] for doc in neighbours[i]] 142 | inferred_labels.append(Counter(labels).most_common(1)[0][0]) 143 | 144 | return inferred_labels, embeddings 145 | 146 | def embed(self, texts): 147 | embeddings = self.embed_model.encode( 148 | texts, 149 | batch_size=self.embed_batch_size, 150 | show_progress_bar=True, 151 | convert_to_numpy=True, 152 | normalize_embeddings=True, 153 | ) 154 | 155 | return embeddings 156 | 157 | def project(self, embeddings): 158 | mapper = UMAP(n_components=self.umap_components, metric=self.umap_metric).fit( 159 | embeddings 160 | ) 161 | return mapper.embedding_, mapper 162 | 163 | def cluster(self, embeddings): 164 | print( 165 | f"Using DBSCAN (eps, nim_samples)=({self.dbscan_eps,}, {self.dbscan_min_samples})" 166 | ) 167 | clustering = DBSCAN( 168 | eps=self.dbscan_eps, 169 | min_samples=self.dbscan_min_samples, 170 | n_jobs=self.dbscan_n_jobs, 171 | ).fit(embeddings) 172 | 173 | return clustering.labels_ 174 | 175 | def build_faiss_index(self, embeddings): 176 | index = faiss.IndexFlatL2(embeddings.shape[1]) 177 | index.add(embeddings) 178 | return index 179 | 180 | def summarize(self, texts, labels): 181 | unique_labels = len(set(labels)) - 1 # exclude the "-1" label 182 | client = InferenceClient(self.summary_model, token=self.summary_model_token) 183 | cluster_summaries = {-1: "None"} 184 | 185 | for label in range(unique_labels): 186 | ids = np.random.choice(self.label2docs[label], self.summary_n_examples) 187 | examples = "\n\n".join( 188 | [ 189 | f"Example {i+1}:\n{texts[_id][:self.summary_chunk_size]}" 190 | for i, _id in enumerate(ids) 191 | ] 192 | ) 193 | 194 | request = self.summary_template.format( 195 | examples=examples, instruction=self.summary_instruction 196 | ) 197 | response = client.text_generation(request) 198 | if label == 0: 199 | print(f"Request:\n{request}") 200 | cluster_summaries[label] = self._postprocess_response(response) 201 | print(f"Number of clusters is {len(cluster_summaries)}") 202 | return cluster_summaries 203 | 204 | def _postprocess_response(self, response): 205 | if self.topic_mode == "multiple_topics": 206 | summary = response.split("\n")[0].split(".")[0].split("(")[0] 207 | summary = ",".join( 208 | [txt for txt in summary.strip().split(",") if len(txt) > 0] 209 | ) 210 | return summary 211 | elif self.topic_mode == "single_topic": 212 | first_line = response.split("\n")[0] 213 | topic, score = None, None 214 | try: 215 | topic = first_line.split("Topic:")[1].split("(")[0].split(",")[0].strip() 216 | except IndexError: 217 | print("No topic found") 218 | try: 219 | score = first_line.split("Educational value rating:")[1].strip().split(".")[0].strip() 220 | except IndexError: 221 | print("No educational score found") 222 | full_output = f"{topic}. Educational score: {score}" 223 | return full_output 224 | else: 225 | raise ValueError( 226 | f"Topic labeling mode {self.topic_mode} is not supported, use single_topic or multiple_topics instead." 227 | ) 228 | 229 | def save(self, folder): 230 | if not os.path.exists(folder): 231 | os.makedirs(folder) 232 | 233 | with open(f"{folder}/embeddings.npy", "wb") as f: 234 | np.save(f, self.embeddings) 235 | 236 | faiss.write_index(self.faiss_index, f"{folder}/faiss.index") 237 | 238 | with open(f"{folder}/projections.npy", "wb") as f: 239 | np.save(f, self.projections) 240 | 241 | with open(f"{folder}/cluster_labels.npy", "wb") as f: 242 | np.save(f, self.cluster_labels) 243 | 244 | with open(f"{folder}/texts.json", "w") as f: 245 | json.dump(self.texts, f) 246 | 247 | with open(f"{folder}/mistral_prompt.txt", "w") as f: 248 | f.write(DEFAULT_INSTRUCTION) 249 | 250 | if self.cluster_summaries is not None: 251 | with open(f"{folder}/cluster_summaries.json", "w") as f: 252 | json.dump(self.cluster_summaries, f) 253 | 254 | def load(self, folder): 255 | if not os.path.exists(folder): 256 | raise ValueError(f"The folder '{folder}' does not exsit.") 257 | 258 | with open(f"{folder}/embeddings.npy", "rb") as f: 259 | self.embeddings = np.load(f) 260 | 261 | self.faiss_index = faiss.read_index(f"{folder}/faiss.index") 262 | 263 | with open(f"{folder}/projections.npy", "rb") as f: 264 | self.projections = np.load(f) 265 | 266 | with open(f"{folder}/cluster_labels.npy", "rb") as f: 267 | self.cluster_labels = np.load(f) 268 | 269 | with open(f"{folder}/texts.json", "r") as f: 270 | self.texts = json.load(f) 271 | 272 | if os.path.exists(f"{folder}/cluster_summaries.json"): 273 | with open(f"{folder}/cluster_summaries.json", "r") as f: 274 | self.cluster_summaries = json.load(f) 275 | keys = list(self.cluster_summaries.keys()) 276 | for key in keys: 277 | self.cluster_summaries[int(key)] = self.cluster_summaries.pop(key) 278 | 279 | # those objects can be inferred and don't need to be saved/loaded 280 | self.id2cluster = { 281 | index: label for index, label in enumerate(self.cluster_labels) 282 | } 283 | self.label2docs = defaultdict(list) 284 | for i, label in enumerate(self.cluster_labels): 285 | self.label2docs[label].append(i) 286 | 287 | self.cluster_centers = {} 288 | for label in self.label2docs.keys(): 289 | x = np.mean([self.projections[doc, 0] for doc in self.label2docs[label]]) 290 | y = np.mean([self.projections[doc, 1] for doc in self.label2docs[label]]) 291 | self.cluster_centers[label] = (x, y) 292 | 293 | def show(self, interactive=False): 294 | df = pd.DataFrame( 295 | data={ 296 | "X": self.projections[:, 0], 297 | "Y": self.projections[:, 1], 298 | "labels": self.cluster_labels, 299 | "content_display": [ 300 | textwrap.fill(txt[:1024], 64) for txt in self.texts 301 | ], 302 | } 303 | ) 304 | 305 | if interactive: 306 | self._show_plotly(df) 307 | else: 308 | self._show_mpl(df) 309 | 310 | def _show_mpl(self, df): 311 | fig, ax = plt.subplots(figsize=(12, 8), dpi=300) 312 | 313 | df["color"] = df["labels"].apply(lambda x: "C0" if x==-1 else f"C{(x%9)+1}") 314 | 315 | df.plot( 316 | kind="scatter", 317 | x="X", 318 | y="Y", 319 | c="labels", 320 | s=0.75, 321 | alpha=0.8, 322 | linewidth=0, 323 | color=df["color"], 324 | ax=ax, 325 | colorbar=False, 326 | ) 327 | 328 | for label in self.cluster_summaries.keys(): 329 | if label == -1: 330 | continue 331 | summary = self.cluster_summaries[label] 332 | position = self.cluster_centers[label] 333 | t= ax.text( 334 | position[0], 335 | position[1], 336 | summary, 337 | horizontalalignment='center', 338 | verticalalignment='center', 339 | fontsize=4, 340 | ) 341 | t.set_bbox(dict(facecolor='white', alpha=0.9, linewidth=0, boxstyle='square,pad=0.1')) 342 | ax.set_axis_off() 343 | 344 | def _show_plotly(self, df): 345 | fig = px.scatter( 346 | df, 347 | x="X", 348 | y="Y", 349 | color="labels", 350 | hover_data={"content_display": True, "X": False, "Y": False}, 351 | width=1600, 352 | height=800, 353 | color_continuous_scale="HSV", 354 | ) 355 | 356 | fig.update_traces(hovertemplate="%{customdata[0]}") 357 | 358 | fig.update_traces( 359 | marker=dict(size=1, opacity=0.8), # color="white" 360 | selector=dict(mode="markers"), 361 | ) 362 | 363 | fig.update_layout( 364 | template="plotly_dark", 365 | ) 366 | 367 | # show cluster summaries 368 | for label in self.cluster_summaries.keys(): 369 | if label == -1: 370 | continue 371 | summary = self.cluster_summaries[label] 372 | position = self.cluster_centers[label] 373 | 374 | fig.add_annotation( 375 | x=position[0], 376 | y=position[1], 377 | text=summary, 378 | showarrow=False, 379 | yshift=0, 380 | ) 381 | 382 | fig.show() 383 | --------------------------------------------------------------------------------