| --- |
| license: other |
| license_name: betterup-candor-terms |
| license_link: LICENSE |
| pretty_name: "CANDOR Corpus (Raw recordings)" |
| language: |
| - en |
| multilinguality: |
| - monolingual |
| size_categories: |
| - 1K<n<10K |
| tags: |
| - conversation |
| - dyadic |
| - multimodal |
| - speech |
| - video |
| - psychology |
| extra_gated_heading: "Request access to the CANDOR corpus (raw)" |
| extra_gated_description: >- |
| Choose the version you need below. Approval also opens the other two versions, |
| so you can switch later without applying again. Requests are reviewed by BetterUp, |
| usually within a few days. Vague or one-line |
| research descriptions are sent back for more detail, so please write 3 to 5 |
| sentences covering your research question, your method, and which parts of the |
| data you need. Two examples of the level of detail we look for: "I want to use |
| this dataset to understand how people start conversations, what strategies they |
| use to establish grounding, and what topics they work with." "My research explores |
| the psycholinguistics of dyadic interaction; I'm looking at the correlation between |
| turn-by-turn language style matching and sentiment." An institutional email is |
| required. BetterUp uses your answers only to review your request and to contact |
| you about CANDOR. |
| extra_gated_prompt: >- |
| Access to the CANDOR Corpus is governed by the CANDOR Corpus License, Version 2.1. |
| Read it before you agree: https://huggingface.co/datasets/betterup/candor-raw#candor-corpus-license In short: you may use the corpus |
| for research, development, and commercial purposes; you may not redistribute it, |
| share derived or synthetic versions of it, identify or contact participants, |
| create biometric identifiers, or create synthetic media of participants; you must |
| cite the corpus and delete anything BetterUp removes from it. The full license text |
| controls. Questions: datarequests@betterup.co. |
| extra_gated_fields: |
| Full name: text |
| Institutional email (not Gmail, Outlook, QQ or similar): text |
| Research institution or organization: text |
| Which version of CANDOR do you need?: |
| type: select |
| options: |
| - "Tabular: transcripts, features, surveys, metadata (about 11 GB)" |
| - "Processed: tabular plus audio and processed video (about 350 GB)" |
| - "Raw: processed plus the original recordings (about 950 GB)" |
| Country: country |
| Describe your intended use in 3 to 5 sentences (question, method, which parts of the data): text |
| Were you approved for CANDOR before November 2026?: |
| type: select |
| options: |
| - "No" |
| - "Yes, with the institutional email above" |
| - "Yes, with a different email" |
| Would you also like access to CANDOR on CMU TalkBank?: |
| type: select |
| options: |
| - "Yes, I want to access CANDOR on CMU TalkBank" |
| - "No, I do not want to access CANDOR on CMU TalkBank" |
| - "I'm not sure / I don't know what TalkBank is" |
| I have read and agree to the CANDOR Corpus License: checkbox |
| extra_gated_button_content: "Agree and send request" |
| viewer: false |
| --- |
| |
| # CANDOR Corpus: Raw |
|
|
| The CANDOR corpus is a large multimodal dataset of naturalistic, unscripted conversations between pairs of strangers, recorded over video call in 2020 and released by BetterUp Labs. It accompanies Reece et al. (2023), "The CANDOR corpus: Insights from a large multimodal dataset of naturalistic conversation," *Science Advances*. |
|
|
| Participants were adult volunteers in the United States, recruited through Prolific, an online research platform, and paid for their time. Each consented to their recordings being made publicly available. The corpus contains no BetterUp customer, member, or coaching data. |
|
|
| The corpus has 1,656 conversations among 1,460 participants, about 860 hours of conversation in total (counting the time both participants were on the call), with a median length of 29 minutes. The 2024 transcription contains about 9.8 million words. |
|
|
| This repository holds the **raw** flavor. The corpus is published in three flavors, each its own gated repository: |
|
|
| | Flavor | Repo | Approx. size | Contents | |
| | --- | --- | --- | --- | |
| | Tabular (no media) | `betterup/candor-tabular` | ~11 GB | Transcripts, turn-level tables, surveys, metadata | |
| | Processed | `betterup/candor-processed` | ~350 GB | Audio and derived per-frame / per-turn features | |
| | Raw | `betterup/candor-raw` | ~950 GB | Original video and audio recordings | |
|
|
| ## Versions |
|
|
| | Version | Date | Summary | |
| | --- | --- | --- | |
| | v1.1 | November 2026 | New 2024 transcription, NaturalTurn transcripts, updated backchannel cues, corrected audio and video features, corrected survey scores | |
| | v1.0 | February 2022 | Initial release | |
|
|
| v1.1 keeps the v1.0 archives' layout and conversation groupings and changes two files inside them: each conversation's `audio_video_features.csv` and `survey.csv` are replaced with the corrected versions described below. Every other file in the archives is byte-for-byte identical to v1.0. The new transcript tables live in `betterup/candor-tabular` under `transcripts/`. The original v1.0 tabular files stay available at the `v1.0` revision of `betterup/candor-tabular`, for example with `hf download betterup/candor-tabular --repo-type dataset --revision v1.0`. |
|
|
| ### What's new in v1.1 |
|
|
| **New speech-to-text pass.** Every conversation was re-transcribed with AWS Transcribe in January 2024. The two passes find nearly the same number of words, but the 2024 pass switches between speakers less often, so turns built from it run longer on average. It also gives a confidence score for every word. The raw 2024 output is in `betterup/candor-tabular` under `transcripts/aws_2024_raw/`. |
|
|
| **NaturalTurn transcripts.** v1.1 adds transcripts built with NaturalTurn ([Cooney & Reece, 2025](https://doi.org/10.1038/s41598-025-24381-1)), a turn model designed to match how people experience conversational turns. NaturalTurn keeps a speaker's turn whole when the listener says "mhm" or "yeah", and records the listener's response alongside the turn it answered. |
|
|
| **v1.0 turn models rebuilt.** Audiophile, Backbiter, and Cliffhanger were regenerated from the 2024 transcription with the same algorithms as v1.0, so all four turn models now come from one transcription pass. All four were produced with [`betterup/natural-turn-transcription`](https://github.com/betterup/natural-turn-transcription) at commit [`d013669`](https://github.com/betterup/natural-turn-transcription/commit/d013669). |
|
|
| **Updated backchannel cues.** Backbiter and NaturalTurn decide whether a short reply is a backchannel partly by checking it against a list of cue words. v1.1 adds "ok" and "hm", common spellings in the 2024 transcription. It drops "a" and "e", which had caused article fragments and spelled-out letters to count as acknowledgments. Audiophile and Cliffhanger do not use the list. |
|
|
| **Corrected audio and video features.** v1.1 replaces every `audio_video_features.csv` with a corrected version. Starting from the v1.0 file, it makes five fixes: |
|
|
| - **Video timing.** The v1.0 video pipeline applied each recording's start offset twice, so video features landed later in the conversation than they happened. The delay equaled the gap between the call starting and that participant's recording starting. For half of participants that gap was 2 seconds or less. For a third it was a minute or more, and it reached 52 minutes for participants who joined late or reconnected. It affected the video features of 2,922 of 3,312 participants, in 1,654 of 1,656 conversations. v1.1 moves each video value back to the second it belongs to. No video values were re-estimated or dropped. |
| - **Sound measurements during padding.** Before a late joiner arrived, and during reconnect gaps, the audio track was padded with silence, and v1.0 recorded sound measurements for that padding. v1.1 blanks the acoustic features for 291,038 participant-seconds (about 1 in 27) in which that participant had no recorded audio. Real silences during a recording are untouched. |
| - **Empty seconds after the call ended.** The timing bug stretched some files past the end of the call. v1.1 removes 702,946 rows in 1,477 conversations that fall after the recording ended and hold no data. Use the new row counts as conversation lengths. |
| - **Invalid energy values.** 708 `log_energy` values held a huge negative placeholder left by empty audio frames (below -10^20) and are now blank. |
| - **Final stretch of speech.** v1.0 missed each speaker's last stretch of speech when setting `is_speaking`. v1.1 restores 5,541 seconds in 1,519 conversations, using the original transcripts and rules. |
|
|
| Analyses that aligned v1.0 video features with speech, transcripts, or the other participant should be rerun. |
|
|
| **Corrected survey scores.** Only 256 of the 3,312 participant-conversations completed four extra questionnaires: the Reading the Mind in the Eyes test, the empathy subscales of the Interpersonal Reactivity Index (perspective taking and empathic concern), two Autism-Spectrum Quotient subscales (social skill and communication), and the Brief Fear of Negative Evaluation scale. v1.0 gave the other 3,056 a score of 0 on all six measures instead of leaving them blank. v1.1 blanks them. For the 256 who did complete the questionnaires, v1.0 scored the empathy subscales with the wrong item coding, so both scores ran low: perspective taking by 8 points and empathic concern by 12. v1.1 recomputes them from the original item responses. No other survey values change. |
|
|
| ### Known limitations |
|
|
| These come from how the 2020 recordings were processed, and v1.1 does not fix them. The lists named below are in `betterup/candor-tabular`, as Parquet and CSV. |
|
|
| - **Video timing is accurate to about one second.** The original frame timestamps no longer exist, so video values were moved by whole seconds using the call metadata. |
| - **Video features stop early in some recordings.** A screen found 75 recordings, in 74 conversations, whose video features end long before the recording does. In the two cases checked by hand the video continued, so these look like extraction failures. The missing values stay blank. `features/quality/video_cutoffs` lists them, and the list may not be complete. |
| - **Some audio ends early.** The 2020 audio used for transcription and acoustic features ends when the shorter of the two participants' streams ends. In 387 conversations, the longer stream continued for more than 10 seconds. In the three cases checked by hand, one participant had already left the call. `features/quality/audio_tail_candidates` lists them. |
| - **`is_speaking` follows the v1.0 transcription, not the 2024 one.** It marks which participant held the turn in each second, including short pauses, so it is not a measure of audible speech. |
| - **First and last turns can include audio from outside the conversation.** See the note below. |
| - **Some seconds have partial audio.** `features/quality/coverage` lists every participant-second with partial or no recorded audio. Use it, rather than row counts alone, when computing per-second rates. |
| |
| #### Note on first and last turns |
| |
| Recording began when the first participant joined the call, not when both participants were present. As a result, the first turn (`turn_id = 0`) can include time before the second participant arrived. During this stretch the transcript may pick up background noise, other voices in the room, audio from a TV or other device, or the participant talking before the conversation starts. These passages can run for several minutes and often read as disconnected fragments. They are not part of the conversation. The last turn can have the same problem when one participant stays on the call after the other leaves. |
| |
| We recommend inspecting the first and last turns of each conversation and trimming or removing them before analysis. An unusually long first or last turn is the clearest sign. Note that the first participant's real opening words can be merged into the same turn as the pre-join audio, so removing turn 0 entirely may occasionally cut a genuine greeting. |
| |
| ## Files |
| |
| | Path | Contents | |
| | --- | --- | |
| | `data/*_media_part_NNN.zip` | 166 zip archives of 6 to 10 conversations each, 3.5 to 7.5 GB apiece, about 952 GB in total | |
| | `manifest.yaml` | md5 checksum, size, and conversation list for every archive in all three flavors | |
| |
| Each archive holds one folder per conversation, named by conversation ID: |
| |
| | Path in the conversation folder | Contents | |
| | --- | --- | |
| | `metadata.json` | Call metadata: each participant's ID, audio channel, and recording files with their start and stop offsets | |
| | `survey.csv` | Pre- and post-conversation survey responses, one row per participant (corrected in v1.1) | |
| | `audio_video_features.csv` | Acoustic and video features, one row per participant per second (corrected in v1.1) | |
| | `transcription/transcript_audiophile.csv`, `transcript_backbiter.csv`, `transcript_cliffhanger.csv` | v1.0 transcripts | |
| | `transcription/transcribe_output.json` | v1.0 speech-to-text output | |
| | `processed/<conversation_id>.mp4`, `.mp3` | Conversation-level video, and stereo audio with one participant per channel | |
| | `processed/<participant_id>.mp4` | Each participant's video | |
| | `processed/channel_map.json`, `thumbnail.png` | Which audio channel (L or R) belongs to which participant, and a preview image | |
| | `raw/<conversation_id>.json` | The video platform's record of the session: each recording file with its start and stop offsets | |
| | `raw/<recording_id>.mkv` | Original recordings as captured, one file per recording segment | |
| |
| The v1.1 transcripts (all four turn models from the 2024 transcription) are tables in `betterup/candor-tabular`, not files in these archives. |
| |
| ## Finding a conversation |
| |
| `manifest.yaml` lists the conversations in each archive. To download only the archive that holds one conversation: |
| |
| ```python |
| import yaml |
| from huggingface_hub import hf_hub_download |
| |
| manifest = yaml.safe_load(open(hf_hub_download("betterup/candor-raw", "manifest.yaml", repo_type="dataset"))) |
| conversation = "0020a0c5-1658-4747-99c1-2839e736b481" |
| archive = next(a["filename"] for v in manifest["media_versions"] for a in v["archives"] |
| if a["filename"].startswith("raw_media") and conversation in a["conversations"]) |
| hf_hub_download("betterup/candor-raw", f"data/{archive}", repo_type="dataset", local_dir=".") |
| ``` |
| |
| Check each archive against the md5 in `manifest.yaml` after download. |
| |
| ## Data dictionary and other documentation |
| |
| These live in `betterup/candor-tabular` and cover all three versions: |
| |
| | Path | Contents | |
| | --- | --- | |
| | `docs/data_dictionary/` | Every file in a conversation folder, every audio and video feature with how it was computed, every transcript column, and every survey question with its exact wording | |
| | `annotations/qualitative_review/` | The research team's review of each conversation (recording problems, where the conversation really starts and ends, notes on trimming), plus their notes on conversations that stood out | |
| | `replication/sciadv-2023/` | Analysis code for the CANDOR corpus paper | |
| | `replication/naturalturn-2025/` | Analysis code and data for the NaturalTurn paper | |
| |
| ## Access |
| |
| Access is granted per person after review. Click **Agree and send request** above and fill in every field. Choose the version you need in the form. Approval also opens the other two versions, so you can switch later without applying again. Approved users download with the Hugging Face CLI: |
| |
| ```bash |
| hf auth login |
| hf download betterup/candor-raw --repo-type dataset --local-dir ./candor-raw |
| ``` |
| |
| Downloads resume if interrupted. Use `--include` to fetch a subset. |
| |
| ## Terms of use |
| |
| The CANDOR Corpus License, Version 2.1, governs all use. Its full text is at the end of this card and in `LICENSE`. In short: research, development, and commercial use are allowed. You may make derived or synthetic versions of the corpus for your own use, but not share them. Redistributing the corpus, identifying participants, and building biometric identifiers are not allowed. Benchmarks that refer to conversations by ID and timestamp are allowed. If BetterUp removes any part of the corpus, it is listed in `REMOVALS.md` in `betterup/candor-tabular`. Within 30 days you must delete your copies of it and drop references to it from any benchmark you maintain. |
| |
| ## Citation |
| |
| Please cite the CANDOR corpus paper. If you use the NaturalTurn transcripts, please also cite the NaturalTurn paper. Replication code for both papers is under `replication/` in `betterup/candor-tabular`. To cite this dataset release itself, use its DOI, [10.57967/hf/11411](https://doi.org/10.57967/hf/11411). |
| |
| ```bibtex |
| @article{reece2023candor, |
| title = {The CANDOR corpus: Insights from a large multimodal dataset of naturalistic conversation}, |
| author = {Reece, Andrew and Cooney, Gus and Bull, Peter and Chung, Christine and Dawson, Bryn and Fitzpatrick, Casey and Glazer, Tamara and Knox, Dean and Liebscher, Alex and Marin, Sebastian}, |
| journal = {Science Advances}, |
| volume = {9}, |
| number = {13}, |
| pages = {eadf3197}, |
| year = {2023}, |
| doi = {10.1126/sciadv.adf3197} |
| } |
| |
| @article{cooney2025naturalturn, |
| title = {NaturalTurn: a method to segment speech into psychologically meaningful conversational turns}, |
| author = {Cooney, Gus and Reece, Andrew}, |
| journal = {Scientific Reports}, |
| volume = {15}, |
| pages = {39155}, |
| year = {2025}, |
| doi = {10.1038/s41598-025-24381-1} |
| } |
| |
| @misc{betterup2026candorraw, |
| title = {CANDOR Corpus (Raw recordings)}, |
| author = {{BetterUp, Inc.}}, |
| year = {2026}, |
| version = {v1.1}, |
| publisher = {Hugging Face}, |
| doi = {10.57967/hf/11411}, |
| url = {https://huggingface.co/datasets/betterup/candor-raw} |
| } |
| ``` |
| |
| ## Contact |
| |
| datarequests@betterup.co |
| |
| ## CANDOR Corpus License |
| |
| **Notice to CANDOR participants:** Consistent with your original submission, you maintain the right to have your data removed from the CANDOR Corpus. You may do so by emailing the email provided in your original consent form, or you may contact datarequests@betterup.co. |
|
|
| **CANDOR Corpus License** |
| **Version 2.1, effective November 3, 2026** |
|
|
| These terms govern every person and organization that requests access to, downloads, or uses any part of the CANDOR Corpus. You accept them when you access, download, or use any part of the CANDOR Corpus. By your acceptance, you represent that you have read and understood these terms. If you accept on behalf of a company or other organization, you also represent that you are authorized to accept these terms on its behalf and to bind it and its affiliates, and "you" includes that organization and its affiliates. If you accepted the earlier CANDOR terms of use, you may keep using your copies under those terms. If you accept these terms, they replace the earlier terms for all of your copies of the CANDOR Corpus. |
|
|
| ### 1. Definitions |
|
|
| - **"Benchmark"** means a benchmark, leaderboard, or other evaluation suite that others can use or compare results against. |
| - **"BetterUp"** means BetterUp, Inc. |
| - **"CANDOR Corpus"** means the CANDOR Corpus of recorded conversations released by BetterUp under these terms, including its video and audio recordings, transcripts, survey responses, extracted features, annotations, and metadata, and any copy or portion of it, original or modified, however obtained. |
| - **"Derivative"** means any dataset, or other collection of data or media, made from the CANDOR Corpus, directly or by using a Model or any other tool, that reproduces, imitates, or is meant to stand in for any part of the CANDOR Corpus. Examples include synthetic variants such as re-voiced or re-synthesized recordings, paraphrased or translated transcripts, and other synthetic data generated to imitate data in the CANDOR Corpus. A Derivative does not include a Model, or anything clause 3(b) allows you to share. |
| - **"Hugging Face"** means the Hugging Face platform at https://huggingface.co, and any successor website or platform. |
| - **"Model"** means a machine learning model, weights, adapters, and checkpoints, whose training or fine-tuning uses any part of the CANDOR Corpus, or uses data generated by another machine learning model, weights, adapters, and checkpoints trained or fine-tuned on the CANDOR Corpus. |
| - **"Participant"** means each person who took part in a recorded conversation in the CANDOR Corpus. |
| - **"Removed Material"** means any part of the CANDOR Corpus that BetterUp has listed as removed in the removals log described in clause 8(a). |
|
|
| ### 2. License grant |
|
|
| BetterUp grants you a worldwide, royalty-free, non-exclusive, non-transferable license to use the CANDOR Corpus for any lawful purpose, including research, development, and commercial purposes, subject to these terms. For clarity, anything within this license that states that something is permitted still means that that permission is subject to the remainder of the terms. |
|
|
| ### 3. Sharing and redistribution |
|
|
| - **(a) General rule.** Except as clause 3(b) or 3(c) allows, you may not redistribute, republish, sublicense, or relicense any part of the CANDOR Corpus, including as part of another dataset, Benchmark, training corpus, or mirror, or by making its content available through any service or system, such as a search or retrieval tool, without BetterUp's written permission. |
| - **(b) Sharing allowed without permission.** You may share the following: |
| - **(i)** references to specific material by conversation identifier and timestamp, and annotations or labels keyed to them, as long as they do not reproduce the material itself. Anyone who wants the material must obtain it from BetterUp by accepting these terms; |
| - **(ii)** short quotations of transcript text in a research paper, preprint, or similar publication, where needed to illustrate a finding. This does not permit sharing audio, video, or images of any Participant; and |
| - **(iii)** aggregate statistics and research findings that do not reproduce the CANDOR Corpus or identify any Participant. |
| - **(c) Models.** You may share a Model, or make it available for others to use, including as a service, if you: |
| - **(i)** require each recipient or user to comply with clauses 3(e), 5, and 6; and |
| - **(ii)** include with the Model, or in its documentation, a notice that it was trained using the CANDOR Corpus from BetterUp Labs, BetterUp, Inc., with a link to the CANDOR Corpus repository used on Hugging Face (https://huggingface.co/datasets/betterup/candor-tabular, https://huggingface.co/datasets/betterup/candor-processed, or https://huggingface.co/datasets/betterup/candor-raw). |
| - **(d) Model outputs.** Outputs generated by a Model are not part of the CANDOR Corpus, unless they reproduce any part of it. A dataset of Model outputs may still be a Derivative, which clause 3(e) governs. |
| - **(e) Derivatives.** You may create and use Derivatives under the license in clause 2. You may not redistribute, republish, sublicense, or relicense a Derivative, or otherwise share it or make it available to anyone other than you, including as part of a Benchmark or through any service or system, without BetterUp's written permission. |
| - **(f)** All sharing under this clause remains subject to clauses 4 through 6. |
|
|
| ### 4. Benchmarks |
|
|
| - **(a)** You may use the CANDOR Corpus to build and publish a Benchmark, as long as it shares the CANDOR Corpus only as clause 3(b) allows, for example by referring to material by conversation identifier and timestamp. Anyone who uses the Benchmark must obtain the material from BetterUp by accepting these terms. |
| - **(b)** Including in a Benchmark any part of the CANDOR Corpus beyond what clause 3(b) allows, or any Derivative, or sharing either for use in a Benchmark, requires BetterUp's written permission under clause 3. To request permission, contact BetterUp at datarequests@betterup.co. Any permission BetterUp grants is subject at all times to clauses 5 through 9 and to any additional conditions BetterUp sets in writing, including attribution on the Benchmark's page under clause 7(b) and removal of Removed Material from the Benchmark under clauses 4(c) and 8. |
| - **(c)** You must remove all references to Removed Material, and any annotations of it, from any Benchmark you maintain, within the period clause 8(b) sets. |
| - **(d)** Subject to the above, you may evaluate any machine learning model (whether or not it is a Model) on the CANDOR Corpus and report the results, including in papers, preprints, and model cards. |
|
|
| ### 5. Participant privacy |
|
|
| Participants' personal privacy in the CANDOR Corpus must be respected, including, inter alia, by you adhering to the following: |
|
|
| - **(a)** You may not use any part of the CANDOR Corpus to identify, locate, contact, or gather information about any Participant beyond what the CANDOR Corpus provides. |
| - **(b)** You may not link any Participant to data outside the CANDOR Corpus. |
| - **(c)** You may not use the CANDOR Corpus to seek out, extract, or compile personal details that a Participant mentions about themselves in a recording. |
| - **(d)** You may not use the recordings to create synthetic media that depicts, imitates, or impersonates any Participant. If you do create synthetic media based in any part on the CANDOR Corpus, you must ensure that any such synthetic media does not depict, imitate, or otherwise bear similar likeness to any Participant. |
| - **(e)** If you share a Model, a Derivative, or any other dataset or work made using the CANDOR Corpus, you must take reasonable measures to prevent it from revealing any Participant's personal details, likeness, or voice. Examples of such measures include testing for and reducing memorization of training data, filtering outputs, removing personal details from data you share, and privacy-preserving training methods such as differential privacy. |
| - **(f)** Training, fine-tuning, or evaluating a machine learning model on recordings or transcripts as a whole does not by itself violate this clause. |
|
|
| ### 6. Biometric identifiers |
|
|
| - **(a)** You may not use the CANDOR Corpus, or any component or element thereof, to create face prints, voiceprints, or other biometric identifiers or templates of any Participant, or to train or test any system that recognizes or verifies a person's identity. |
| - **(b)** General modeling of facial expression, speech, voice characteristics, and behavior is permitted, as long as it does not identify any Participant. This includes temporarily telling speakers apart within a single recording (for example, speaker diarization), provided the result is not kept or used to identify anyone. |
|
|
| ### 7. Attribution |
|
|
| - **(a)** Any publication, preprint, report, or other public work that uses the CANDOR Corpus must cite: |
| - Reece, A., et al. (2023). The CANDOR corpus: Insights from a large multimodal dataset of naturalistic conversation. *Science Advances*, 9(13), eadf3197; and |
| - if the work uses the NaturalTurn transcripts, Cooney, G., & Reece, A. (2025). NaturalTurn: A method to segment speech into psychologically meaningful conversational turns. *Scientific Reports*, 15, 39155. |
| - **(b)** Any public model card, dataset card, annotation set, Benchmark page, or product documentation for a work that uses the CANDOR Corpus, or a Model, must state that the work uses the CANDOR Corpus from BetterUp Labs, BetterUp, Inc., with a link to the CANDOR Corpus repository used on Hugging Face (https://huggingface.co/datasets/betterup/candor-tabular, https://huggingface.co/datasets/betterup/candor-processed, or https://huggingface.co/datasets/betterup/candor-raw). |
| - **(c)** Any model or dataset you publish on Hugging Face or any similar platform that was trained or evaluated on, or made using, the CANDOR Corpus must list the CANDOR Corpus repository it used (`betterup/candor-tabular`, `betterup/candor-processed`, or `betterup/candor-raw`) in the `datasets` field of its Hugging Face card metadata, or in the equivalent field on any similar platform. If a platform has no equivalent field, you must list the CANDOR Corpus in the work's description or documentation on that platform. BetterUp requests, but does not require, that any model or dataset you publish on another platform also be published on Hugging Face. |
| - **(d)** These terms do not grant any right to use BetterUp's names or logos except as required by this clause. You may not suggest that BetterUp endorses you or your work. |
| - **(e)** BetterUp requests, but does not require, that you send a link to any public work that uses the CANDOR Corpus to datarequests@betterup.co. |
|
|
| ### 8. Removals |
|
|
| - **(a)** BetterUp may remove any part of the CANDOR Corpus, such as a whole conversation, a single recording, or part of a transcript, including at a Participant's request. Removing a conversation removes its recordings, transcripts, and all other data associated with it. BetterUp lists each removal in a removals log in the CANDOR Corpus repository on Hugging Face at https://huggingface.co/datasets/betterup/candor-tabular and may notify you at the email address used to request access. |
| - **(b)** Within 30 days after BetterUp lists a removal in the removals log, you must delete all copies of the Removed Material, original or modified, and stop using it for any new training, evaluation, or analysis. |
|
|
| ### 9. Compliance with law and indemnity |
|
|
| - **(a)** You are responsible for complying with all laws that apply to your use of the CANDOR Corpus, including privacy, biometric, and AI laws. |
| - **(b)** You will indemnify, defend, and hold harmless BetterUp and its officers, directors, employees, and affiliates against any third-party claims, actions, or proceedings arising out of or in connection with, and to the extent caused by, your use of the CANDOR Corpus in violation of these terms or applicable law. |
|
|
| ### 10. Versions |
|
|
| BetterUp may publish new versions of these terms for future releases of the CANDOR Corpus. If you choose to use a release offered under a new version, that version applies to your use of that release. For the avoidance of doubt, removals under clause 8 do not depend on new versions of these terms: BetterUp may remove any part of the CANDOR Corpus at any time without publishing a new version or a new release, and each removal applies to all copies of the CANDOR Corpus you hold. |
|
|
| ### 11. Termination |
|
|
| Your license terminates automatically if you breach any of these terms, or if you choose to end it. Upon termination, you must stop using the CANDOR Corpus and delete all copies of it that you hold. Models you trained before termination are not affected, unless your license terminated because of your breach. Clauses 3 through 7, clause 9, this clause, and clauses 12 and 13 survive termination. |
|
|
| ### 12. Disclaimer of warranties |
|
|
| THE CANDOR CORPUS IS PROVIDED "AS IS." BETTERUP MAKES NO WARRANTIES, EXPRESS OR IMPLIED, INCLUDING WARRANTIES OF MERCHANTABILITY, TITLE, FITNESS FOR A PARTICULAR PURPOSE, OR NONINFRINGEMENT. BETTERUP MAKES NO WARRANTY THAT ANY PORTION OF THE CANDOR CORPUS IS ACCURATE, COMPLETE, OR ERROR-FREE. TRANSCRIPTS AND FEATURES MAY HAVE BEEN PRODUCED BY AUTOMATED TOOLS AND MAY CONTAIN ERRORS. |
|
|
| ### 13. General |
|
|
| - **(a) Governing law and venue.** These terms are governed by the laws of the State of Delaware, without regard to its conflict of law provisions. The state or federal courts in New Castle County, Delaware will have exclusive jurisdiction over any suit relating to these terms. |
| - **(b) No indirect damages.** TO THE MAXIMUM EXTENT PERMITTED BY LAW, BETTERUP WILL NOT BE LIABLE FOR ANY INDIRECT, SPECIAL, INCIDENTAL, EXEMPLARY, PUNITIVE, OR CONSEQUENTIAL DAMAGES ARISING OUT OF OR IN CONNECTION WITH THESE TERMS OR THE CANDOR CORPUS, REGARDLESS OF THE THEORY OF LIABILITY, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGES. |
| - **(c) Liability cap.** TO THE MAXIMUM EXTENT PERMITTED BY LAW, BETTERUP'S TOTAL LIABILITY UNDER THESE TERMS WILL NOT EXCEED $100. |
| - **(d) Entire agreement.** These terms are the entire agreement between you and BetterUp about the CANDOR Corpus, unless you and BetterUp sign a separate written agreement. |
| - **(e) Contact.** Permission requests and other questions: datarequests@betterup.co. |
|
|
