juliehunter's picture
Update README.md
12a8df2 verified
metadata
pretty_name: Claire English Dialogue Dataset (CEDD)
license: cc-by-nc-sa-4.0
language:
  - en
multilinguality:
  - monolingual
size_categories:
  - 100M<n<1B
task_categories:
  - text-generation
  - text2text-generation
task_ids:
  - language-modeling
  - dialogue-modeling
  - dialogue-generation
tags:
  - conversational
  - text-generation
  - conditional-text-generation
  - dialogue-modeling
  - dialogue-generation
viewer: true
configs:
  - config_name: default
    sample_by: paragraph
    data_files:
      - split: train
        path: EN/*/train.txt
      - split: test
        path: EN/*/test.txt

Claire English Dialogue Dataset (CEDD)
A collection of English dialogue transcripts

This is the first packaged version of the datasets used to train the english variants of the Claire family of large language models (OpenLLM-France/Claire-7B-EN-0.1). (A related French dataset can be found here.)

The Claire English Dialogue Dataset (CEDD) is a collection of transcripts of English dialogues from various sources, including parliamentary proceedings, interviews, broadcast, meetings, and free conversations. Each dialogue is split into speech turns, and each speech turn is labeled with the name of the speaker, or a unique identifier if the speaker is unknown.

Dataset composition

CEDD can be broken down into:

  • 962,550 conversations in total (812,705 in train, 11,992 in test)
  • 20,863,917 speech turns in total (18,576,327 in train, 359,527 in test)
  • around 864M words

It is a collection of several independent datasets, classified by the types of conversations they contain. This categorization is designed to more evenly balance the influence of different styles of dialogue on model training and to facilitate future applications of CEDD for which certain types of dialogue might be more helpful than others.

For more information, you can look at the following documents:

  • EN/metadata.csv contains further statistics on the different subcorpora (broken down by train/test splits).

Data sources

Dataset Description Words Turns Conversations License (and conditions)

Parliamentary Proceedings

Europarl The Europarl parallel corpus 56M 214K 11K No copyright restrictions. If you use this data in your research, please contact phi@jhu.edu

Spoken Dialogue

Charlotte Narratives The Charlotte Narrative and Conversation Collection (CNCC) contains 95 narratives, conversations and interviews representative of the residents of Mecklenburg County, North Carolina and surrounding North Carolina communities. 200K 2.7K 93 Available for download and use for research and development, including commercial development
Switchboard The corpus consists of approximately 260 hours of speech and was originally collected by Texas Instruments in 1990-1, under DARPA sponsorship. 3M 290K 2320 LDC User Ageement for Non-Members

Broadcast

MediaSum (GitHub) MediaSum dataset for summarization. A collection of transcripts of CNN and NPR interviews with short summaries. 720M 13M 458K For research purposes only

Meetings

AMI (project page) The AMI Meeting Corpus is a multi-modal data set consisting of 100 hours of meeting recordings. 712K 75K 139 CC BY 4.0
ICSI (project page) About 70 hours of meeting recordings. 804K 64K <1K CC BY 4.0

Assistance

ReDial (GitHub) ReDial (Recommendation Dialogues) is an annotated dataset of dialogues, where users recommend movies to each other. 1.5M 139K 11K CC BY 4.0
OpenDialKG (GitHub) OpenDialKG is a dataset of conversations between two crowdsourcing agents engaging in a dialog about a given topic. 1M 84K 12K CC-BY-NC-4.0
ABCD (GitHub) Action-Based Conversations Dataset. 1.5M 142K 10K MIT
AirDialogue (GitHub) AirDialogue is a benchmark dataset for goal-oriented dialogue generation research. 37M 4.6M 361K Apache License 2.0
MULTIWOZ2_2 (pfb30) Multi-Domain Wizard-of-Oz dataset (MultiWOZ), a fully-labeled collection of human-human written conversations spanning over multiple domains and topics. 1.9M 143K 10.4K Apache License 2.0
MulDoGO2 (GitHub) Conversations from the airline, fastfood, finance, insurance, media, and software domains. 10M 892K 63K CDLA Permissive License

Free Chat

Chit-Chat (GitHub) Open-domain conversational dataset from the BYU Perception, Control & Cognition lab's Chit-Chat Challenge. 2.3M 7.1K 258K MIT License
DailyDialog High-quality multi-turn dialog dataset. 1.2M 102K 13K CC BY-NC-SA 4.0

Misc

British National Corpus (BNC) Collection of samples of written and spoken language from a wide range of sources, designed to represent a wide cross-section of British English, both spoken and written, from the late twentieth century. 110M 663K 0.9K BCN License

Example use (python)

In the following sample_by="paragraph" is important to ensure that each sample corresponds to a full conversation (not just a speech turn).

Load dataset from HuggingFace cache (downloaded under ~/.cache/huggingface/datasets):

from datasets import load_dataset

dataset = load_dataset("OpenLLM-France/Claire-Dialogue-English-0.1", sample_by="paragraph", streaming=True)

Load dataset from raw text files:

from datasets import load_dataset
import glob

path = "path/to/dataset"
train_files = glob.glob(path + "/*/train.txt")
test_files = glob.glob(path + "/*/test.txt")

dataset = load_dataset("text", data_files={"train": train_files, "test": test_files}, sample_by="paragraph", streaming=True)

Iterate on the dataset:

for sample in dataset["train"]:
    train_conversation = sample["text"]
    ...

for sample in dataset["test"]:
    test_conversation = sample["text"]
    ...

Important notes

All datasets were normalized in text files so that:

  • Conversations are separated by a single blank line.
  • Each line corresponds to a single speech turn.
  • Each line begins with a speaker label of the form "[***:]".
  • When speaker names are anonymized or otherwise unknown, speakers are distinguished by numbers in the following format: "[speaker001:]", "[speaker002:]", …
    Otherwise, speakers are labeled with their names or roles, e.g. "[Paul:]", "[John King:]", "[White House Correspondent:]".
  • There are no parentheses: special annotations are always between square brackets.
  • Commong tags include:
    • "[PII]": Personally Identifiable Information (anonymized name...)
    • "[NOISE]": distinct ambient noises
    • "[LAUGHTER]": laughter
  • Depending on the data source, data may or may not include punctuation marks and upper case letters.
  • The data were normalized in various ways including unicode NFC normalization, conversion of unbreakable spaces to spaces, and standardization of punctuation marks ( -> ..., «/»//// -> ").

License

Given that some of the corpora used for training are only available under CC-BY-NC-SA licenses, Claire-Dialogue-English-0.1 is made available under the CC-BY-NC-SA 4.0 license.

Citations

When using the CEDD corpus, please cite this page:

@misc{openllm2024claire_en,
  author = {Julie Hunter and Jérôme Louradour and Virgile Rennard and Ismaïl Harrando and Guokan Shang and Jean-Pierre Lorré},
  title = {The Claire English Dialogue Dataset},
  year = {2024},
  publisher = {HuggingFace},
  journal = {HuggingFace},
  howpublished = {\url{https://huggingface.co/datasets/OpenLLM-France/Claire-Dialogue-English-0.1}},
}

You should also provide citations for all of the original corpora. They are listed below.

Our versions of MediaSum, ReDial, OpenDialKG, ABCD, AirDialogue, MultiWOZ2.2, MulDoGo2 and Chit-Chat were collected from the DialogStudio compilation, which is also to be cited if using these datasets:

Contact

contact@openllm-france.fr