Blame for SOURCES.md

# Source attribution and terms

This release reformats existing human-authored conversations into participant
perspectives. The changes are role assignment, group-preserving train/validation/test
splits, explicit source instruction extraction, and the source-specific filtering
listed below. Target utterance wording is retained, except that DeliData export
quoting is decoded: matching outer single quotes are removed and doubled apostrophes
within those wrapped strings become single apostrophes. INSPIRED also decodes its
literal QUOTATION_MARK token to a double quote. Other retained text is unchanged.
Human-authored scripted material
is included; AI messages can appear as context but are not training targets.

Each source retains its applicable terms. No blanket license is assigned by this
collection. The source pages below provide the original releases, credits, and
citation details. License identifiers are given only where explicitly documented.

| Dataset value (prefix `convokit/`) | Source and attribution | Terms / processing |
| --- | --- | --- |
| `contextual-abuse` | [Contextual Abuse Dataset](https://convokit.cornell.edu/documentation/cad.html), Vidgen et al.; ConvoKit conversion by Hao Wan | [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). Eligible human targets and complete recoverable reply paths only. |
| `deli-corpus` | [DeliData](https://convokit.cornell.edu/documentation/deli.html), Karadzhov, Stafford, and Vlachos | Original distribution's terms apply; no additional license is inferred. MESSAGE original text retained with the quote decoding described above. |
| `diplomacy-corpus` | [Deception in Diplomacy](https://convokit.cornell.edu/documentation/diplomacy.html), Peskov et al. | Original source terms apply; no additional license identifier asserted. Native game splits retained. |
| `fomc-corpus` | [FOMC transcripts](https://convokit.cornell.edu/documentation/fomc.html), Federal Reserve; corpus work by Chenhao Tan and Lillian Lee | Original source terms apply; no additional license identifier asserted. Available speeches ordered by speech index. |
| `gap-corpus` | [Group Affect and Performance](https://convokit.cornell.edu/documentation/gap.html), University of the Fraser Valley; ConvoKit contributors | [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/). Complete group discussions retained. |
| `iq2-corpus` | [Intelligence Squared Debates](https://convokit.cornell.edu/documentation/iq2.html), Zhang et al.; ConvoKit conversion by Lucas Van Bramer and Marianne Aubin Le Quere | Original source terms apply; no additional license identifier asserted. Collective audience is context only. |
| `llm-facilitation-corpus` | [ConvoKit release](https://zissou.infosci.cornell.edu/convokit/datasets/llm-facilitation-corpus/); Bringing Everyone to the Table, Microsoft Research/GRAIL | Original release terms apply; the downloaded corpus has no standalone license declaration, so none is inferred. Only source-valid conversations; AI facilitators are context only. |
| `news-interview` | [NewsInterview](https://convokit.cornell.edu/documentation/interview.html), original authors and ConvoKit contributors | [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/), as stated in ConvoKit documentation. Human interviews retained. |
| `spolin-corpus` | [SPOLIN](https://convokit.cornell.edu/documentation/spolin.html), Hyundong Justin Cho and Jonathan May | [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/). Both yes-and and non-yes-and human-authored pairs retained; native validation preserved. |

Habermas, CANDOR Cliffhanger, Wikipedia Politeness, and Stack Exchange Politeness
are not included. CANDOR Cliffhanger was removed because of transcript and
segmentation quality.
The two politeness corpora were removed because they lack conversational context.
No personas were generated. This collection does not
claim that its conversation/group splits are disjoint by individual participant.

## Topical-Chat

Copyright Amazon.com, Inc. or its affiliates. Original release:
[alexa/Topical-Chat](https://github.com/alexa/Topical-Chat), revision
`7c939229cbcf6f55f6977b341a5a2f2fe982d53f`.
Topical-Chat and these reformatted Topical-Chat files are provided under the
[Community Data License Agreement – Sharing, Version 1.0](https://github.com/alexa/Topical-Chat/blob/7c939229cbcf6f55f6977b341a5a2f2fe982d53f/DATALICENSE).
This source-specific license does not replace the terms of other sources.

Changes: each conversation is exported from both human participants' perspectives;
roles are reassigned so the target human is `assistant`. Frequent and rare native
validation/test partitions are combined into validation/test respectively.
Dialogue wording and order are unchanged. Reading material and ratings are not
included in training messages; no system messages or generated personas are added.
Original split identities and annotations are retained locally.

## CCPE-M

Original release: [Filip Radlinski, Krisztian Balog, Bill Byrne and Karthik Krishnamoorthi, Google LLC](https://github.com/google-research-datasets/ccpe), revision
`2c9cd30f33f3a154b5a27d015333679262ff36f5`.
Terms: [CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/). These source-specific terms also apply to the reformatted files.

Changes: User-target role reassignment and deterministic conversation-level splits; original dialogue text and order unchanged; annotations omitted.

## CaSiNo

Original release: [Kushal Chawla, Jaysa Ramirez, Rene Clever, Gale Lucas, Jonathan May and Jonathan Gratch](https://github.com/kushalchawla/CaSiNo), revision
`2f6ed4a6a55110152a7699fbaa8150d6036314be`.
Terms: [CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/). These source-specific terms also apply to the reformatted files.

Changes: Both target-human perspectives; interface actions removed without replacement; own resource priorities and reasons formatted as system context; native splits and remaining dialogue text preserved; other metadata omitted.

## INSPIRED

Original release: [Shirley Anugrah Hayati, Dongyeop Kang, Qingxiaoyang Zhu, Weiyan Shi and Zhou Yu](https://github.com/sweetpeach/Inspired), revision
`782d7c1845d11f75d2542809c7f2a51b63ef0304`.
The source repository does not specify a redistribution license. No license or
independently verified redistribution permission is inferred by this collection.

Changes: Seeker-target role reassignment; literal QUOTATION_MARK decoded to double quotes; other text, utterance boundaries, and native splits unchanged; annotations and surveys omitted.

The released files contain 801 training, 99 development, and 99 test conversations,
despite the source documentation describing 100 development and 100 test conversations.
The available conversations are retained without fabricating missing records.

## MutualFriends

Original release: [Stanford NLP's MutualFriends](https://stanfordnlp.github.io/cocoa/),
with the native human–human training conversations distributed by
[ParlAI's pinned loader](https://github.com/facebookresearch/ParlAI/blob/a29567f7ce76992fd1f03c51ba9e3b155a37ea51/parlai/tasks/mutualfriends/agents.py).
The downloaded data file does not include a dataset-specific redistribution
license declaration. No blanket license or independently verified redistribution
permission is inferred from the software license.

Changes: selections and blank messages removed; conversations with fewer than two
speech messages or only one speaking participant excluded. Two human perspectives
are created, with only each target's own friends table and the task goal in system
context. Speech wording, boundaries and order are unchanged. New approximately
90/5/5 train/validation/test splits keep scenarios together within the originally
training-only release. No generated personas or action paraphrases are added.
Some retained speech refers to removed selection events; this conversion does not
claim every exchange is self-contained without the original interface.

## Deal or No Deal

Original release: [Facebook AI Research, end-to-end-negotiator](https://github.com/facebookresearch/end-to-end-negotiator/tree/bbb93bbf00f69fced75d5c0d22e855bda07c9b78).
The source LICENSE is [CC BY-NC 4.0](https://github.com/facebookresearch/end-to-end-negotiator/blob/bbb93bbf00f69fced75d5c0d22e855bda07c9b78/LICENSE).
These source-specific terms apply to the reformatted files and do not change
licenses of other datasets in this collection.

Changes: preserve the existing participant perspectives once each; target speech
becomes assistant, other speech becomes user. Formatting markers and selection
actions are removed. The target's own item quantities and per-item utilities
are formatted as system context; opponent utilities and final allocations are
not included. Released speech strings and official splits are preserved, with
val named validation. Single-utterance records are excluded; valid unmatched
perspectives and failed negotiations remain. No missing counterpart is generated.

The downloaded release has 12,234 perspective records. Removing 64 single-utterance
records leaves 12,170, including 338 unmatched perspectives. Counterpart groups
are inferred from ordered speech and aligned goals; original dialogue IDs are
not supplied, so exported IDs identify native split files and line numbers.

## Taskmaster-4

Original release: [Taskmaster-4, Bill Byrne and Saravanan Ganesh, Google LLC](https://github.com/google-research-datasets/Taskmaster/tree/d92cb6af3005f1dc09c39e75e7daf4a04905e00b/TM-4-2024).
The source README specifies [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
These source-specific terms apply to the reformatted files and do not change
the terms of other datasets in this collection.

Changes: export one customer perspective per conversation; original customer
speech becomes assistant and coffee-agent speech becomes user. Spoken wording,
boundaries, order, short exchanges, and unanswered customer endings are retained.
Collection templates and API annotations are omitted; no system messages or
personas are added. These are human-authored self-dialogues, with some fixed
agent replies used only as context. The separate reward examples are not included.

The eight downloaded files contain 3,710 conversations, versus the README's
claim of 6,500. Retaining the smallest source ID in each complete-dialogue
duplicate group removes 23 extra copies and leaves 3,687 customer perspectives.
Deterministic conversation-level splits contain 3,319 train, 184 validation,
and 184 test rows. These are newly created holdouts; templates can appear in
multiple splits. Raw source identities and duplicate mappings remain local.

## MSC

Original release: [Multi-Session Chat, Jing Xu, Arthur Szlam and Jason Weston,
Facebook AI Research](https://parl.ai/projects/msc/). Retrieved using the ParlAI
release pinned to `a29567f7ce76992fd1f03c51ba9e3b155a37ea51`.

Changes: two human perspectives for eligible native sessions 2–5. Each target's
original persona and supplied previous sessions are labeled system context;
only current-session speech is a training target. Session 1 is not imported as
additional training rows. Explicit speaker identities, wording, chronology and
native chain holdouts are preserved. Malformed speech and unresolved context
conflicts are withheld. Personas are not generated; summaries and opponent
personas are not included.

## FITS

Original release: [Feedback for Interactive Talk & Search, Jing Xu, Megan Ung,
Mojtaba Komeili, Kushal Arora, Y-Lan Boureau and Jason Weston,
Meta AI](https://parl.ai/projects/fits/). Retrieved using the same pinned ParlAI
release. Only v2 and unseen-task test conversations are considered; the earlier
release is not concatenated.

Changes: human requests and corrections become assistant messages, while bot
responses remain user context. Original task instructions and success requirements
are system context. Two documented bot control markers are removed. Search
documents, outcome labels and separately stored correction fields are omitted.
Reviewed self-contained feedback is retained; missing-context and unresolved
review cases remain outside this publication. This is a conservative subset,
not the entire FITS release. Native splits and unseen-task test assignments remain.

For MSC and FITS, the downloaded data files and inspected project pages do not
provide a separate data-license declaration. The ParlAI software license is not
asserted to license these datasets. Attribution is retained without assigning an
invented or blanket dataset license; applicable source terms remain unchanged.

## Humanual-Book

Source: [Humanual-Book, Shirley Wu and colleagues, HumanLM / snap-stanford](https://huggingface.co/datasets/snap-stanford/humanual-book/tree/8272ed5890ae2f226d563fe5edccd14847da0653),
derived from Amazon Reviews 2023. The source declares
[CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/); these
source-specific terms apply to this subset, without assigning a blanket license
to the collection. Credit: Wu, Choi, Khatua, Wang, He-Yueya, Weerasooriya, Wei,
Yang, Leskovec, and Zou, *HUMANLM: Simulating Users with State Alignment Beats
Response Imitation* (2026).

Each row contains a supplied persona as system context, a book title and ordered
description paragraphs as user context, and the source review title and text as
the assistant target. These are review/preference examples, not multi-turn chats.
Ratings, reviewer metadata, and images are not rendered. Supplied personas are
inferred upstream and preserved unchanged; no new personas are generated.

Changes: unpack book descriptions; convert HTML line/paragraph boundaries to
newlines; decode entities once and remove recognized formatting tags and
stylesheet-only HTML comments. Preserve
wording, spelling, punctuation, and literal angle-bracket text. The earliest
review is retained in each of two confirmed duplicate pairs in training. Native
splits are preserved, renaming val to validation: 34,168 train, 492 validation,
and 1,960 test examples. Original IDs and all cleanup/duplicate decisions are
retained in local audits.

The supplied personas may overlap with target reviews; this has not been ruled
out. Reviewers also occur across native splits. These holdouts must not be
described as verified leakage-free or as unseen-reviewer evaluations. Automated
consistency checks confirm source transformation, not authorship or semantic
quality of every individual review.

Sampled source-quality finding: one test example for product `B09M7MR24D` has
a notebook title but a review about a pillow cover. The source product IDs agree;
the original pairing is retained and flagged locally rather than guessed away.
Literal source typos and expressive angle-bracket labels are also preserved.

## Humanual-News, Opinion, Politics, Chat and Email

Original releases from [HumanLM / snap-stanford](https://humanlm.stanford.edu/), not OdysSim copies. Credit: Wu, Choi, Khatua, Wang, He-Yueya, Weerasooriya, Wei, Yang, Leskovec and Zou, *HUMANLM: Simulating Users with State Alignment Beats Response Imitation* (2026).

| Subset | Pinned source | Train | Validation | Test | Source terms |
| --- | --- | ---: | ---: | ---: | --- |
| Humanual-News | [source](https://huggingface.co/datasets/snap-stanford/humanual-news/tree/18f69182151c3591020ddbba76845a336d35ef12) | 45,587 | 706 | 2,270 | Not specified in the pinned dataset card |
| Humanual-Opinion | [source](https://huggingface.co/datasets/snap-stanford/humanual-opinion/tree/26afa1a5cfbc2d4318be56ffe0851c29e1b38b81) | 37,791 | 1,177 | 3,364 | [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/) |
| Humanual-Politics | [source](https://huggingface.co/datasets/snap-stanford/humanual-politics/tree/f4a289dd8d2163f7a9759770344b464664d8c1e0) | 45,425 | 489 | 1,987 | [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/) |
| Humanual-Chat | [source](https://huggingface.co/datasets/snap-stanford/humanual-chat/tree/25ba779f7f9e35273b70c8f883b9fa655cc236d4) | 23,085 | 477 | 1,096 | [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/) |
| Humanual-Email | [source](https://huggingface.co/datasets/snap-stanford/humanual-email/tree/895c918ef386229106ba8689e78910e71cb9548b) | 6,217 | 129 | 514 | [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/) |

Underlying sources are BBC/CNN YouTube posts and comments (News), Reddit discussions
(Opinion), Medium articles and replies (Politics), WildChat human–AI exchanges
(Chat), and Enron email threads (Email). Source-specific and underlying-source
terms apply; no blanket license is assigned to this collection or inferred for News.

Each row preserves one human reply as its sole assistant target. Opinion uses
the original comment in metadata.body, restoring 2,168 source completions changed
by upstream preprocessing (including 1,773 with quote-elision markers). The other
four subsets retain their released completion text. Example identities and
official split membership are unchanged by this restoration. Earlier prompt
history remains as released, including any upstream quote-elision markers in
context; restoration applies to target comments only.
The supplied persona remains system context; prior messages appear as a labeled
user-context block. Previous human speech and AI replies are context only. Text,
code, punctuation, whitespace and message order are preserved without HTML cleanup.
News titles and descriptions and Email subjects come from source metadata.
Empty News descriptions use their supplied titles; missing Email subjects remain
absent. No new personas, missing dialogue, or media context are generated.

Confirmed duplicate replies are kept once, preferring test, then validation,
then train. Blank targets are excluded; conflicting native target versions or
personas are quarantined locally. Identical Email threads attributed to different users are also quarantined.
In other sources, identical wording from distinct identified authors is retained
with an audit. Original identities, split assignments,
speaker mappings, exclusions and duplicate decisions are kept locally.

All retained examples keep their native split; val is named validation. News's
additional full partition is not imported. Chat uses **turn holdouts**, with
earlier and later targets from the same conversation across splits. This is not
an unseen-conversation evaluation. Chat history labels are hashed IPs, which can
change within a released conversation; distinct labels are preserved without
claiming they identify distinct people or the same person. Where anonymized
target IDs cannot be linked to historical authors, neutral labels preserve the
original speaker structure without guessing.

These are supplementary per-reply user-behavior examples, not necessarily complete
chats. Personas are supplied upstream and often inferred. Persona/target overlap,
individual human authorship and semantic coherence of every record have not been
independently established. Links, images, videos and attachments may refer to
material absent from the released text. Native splits are not claimed to be
verified leakage-free. Automated source comparison and sampled review validate
the conversion within these limits.

## LMSYS-Chat-1M (OdysSim subset)

Credit: Lianmin Zheng, Wei-Lin Chiang, Ying Sheng et al.; LMSYS-Chat-1M (2023).
[Original dataset](https://huggingface.co/datasets/lmsys/lmsys-chat-1m)
and [paper](https://arxiv.org/abs/2309.11998).
Prepared source: [OdysSim split release](https://huggingface.co/datasets/Xuhui/sft_processed_large_split/tree/ee332b10508c0a6fbaa86d5de729c89d34f68964/lmsys),
revision `ee332b10508c0a6fbaa86d5de729c89d34f68964`. This configuration contains the processed OdysSim
subset, not the complete million-conversation LMSYS release.

Each retained conversation has one human-target perspective. Original human
`user` messages become `assistant`; original AI `assistant` messages become
`user`. System context, wording, whitespace, ordering, message boundaries,
consecutive messages and unanswered human endings are unchanged. No persona
generation, topic/language filtering, dialogue rewriting or missing-turn
reconstruction was performed. The three approved shorter training prefixes
were removed while preserving their longer versions and those versions' personas.

Train: **79,926**; validation: **401**; test: **170**; total: **80,497**.
These are OdysSim-created splits, with `val` renamed to `validation` and no
retained conversation reassigned. Supplied personas may be inferred task
summaries and may expose later dialogue. They are preserved as supplied, not
verified pre-conversation self-descriptions. Profile fingerprints are not
authenticated human identities; no guarantee of person-disjoint evaluation
or general benchmark decontamination is made.

The original [LMSYS dataset license agreement](https://huggingface.co/datasets/lmsys/lmsys-chat-1m#lmsys-chat-1m-dataset-license-agreement)
contains custom use and redistribution conditions. Separate redistribution
permission was confirmed by the collection owner for this publication.
OdysSim's collection-level Apache-2.0 metadata is not used to relicense the
underlying LMSYS conversations. No blanket license is assigned here.

Complete normalized dialogues and substantive full-prefix overlaps were checked
against the existing published collection. This does not establish the absence
of interior shared passages, semantic duplicates or external benchmark items.
Anonymization placeholders and original AI errors remain in context. Every
exported message was compared with source; sampled semantic review cannot
establish the quality or human authorship of every individual record.

## StudyChat (OdysSim processed histories)

Credit: Hunter McNichols, Fareya Ikram and Andrew Lan, *The StudyChat Dataset:
Analyzing Student Dialogues With ChatGPT in an Artificial Intelligence Course*,
LAK 2026. [Original dataset](https://huggingface.co/datasets/wmcnicho/StudyChat)
and [paper](https://arxiv.org/abs/2503.07928).
The original source documents consenting human students interacting with an AI
assistant in UMass COMPSCI 383 and declares **CC BY 4.0**. Preserve the original
source's terms; OdysSim's collection metadata does not relicense these records.
No blanket license is assigned to this combined collection.

Prepared source: [OdysSim split release](https://huggingface.co/datasets/Xuhui/sft_processed_large_split/tree/ee332b10508c0a6fbaa86d5de729c89d34f68964/studychat),
revision `ee332b10508c0a6fbaa86d5de729c89d34f68964`. Original-source verification used revision
`24d7987d9fbb30d9da12acc53455a10f1cdd2d7f`. All 2,214 supplied histories match an original snapshot
after comparison-only outer whitespace trimming and omission of blank messages.
Original interaction rows and complete conversation histories have different
counts; no original-source interaction rows were additionally imported.

Human student `user` messages become training-target `assistant` messages;
AI tutor `assistant` messages become `user` context. Supplied system personas,
retained text, order, message boundaries, consecutive speakers, and final
student messages remain unchanged. No personas or dialogue are generated.
The 11 tutor-first histories were verified against original history snapshots
and retained without reordering. Source-recorded order is not an independent
verification of real-world chronology. References to local code, files, or
assignments and coherent requests for clarification remain in the dialogue;
no unavailable material was fabricated or appended.

Excluded 478 histories containing only one student message (440 train, 10
validation, 28 test), and six shorter training prefixes in favor of their
longer conversations. Retained personas are each longer record's own context.
Counts: **1,523 train; 20 validation; 187 test; 1,730 total**.
OdysSim assignments are unchanged, with `val` renamed to `validation`.

Supplied personas are inferred context and may describe later dialogue. They
are not verified prior student instructions. The splits share student IDs and
are **not student-disjoint or verified leakage-free evaluations**. Original
AI errors remain as context; this is a human-behavior training source rather
than a set of verified tutoring answers.

Full normalized dialogues and substantive full-prefix overlaps were checked
within the retained source and against the published collection. This does
not establish the absence of arbitrary semantic or interior-subsequence overlap.
Every exported message was compared with its OdysSim source. Seeded samples,
all 11 tutor-first histories, and all six prefix pairs were reviewed; sampled
semantic checks do not establish the quality of every individual utterance.

## DailyDialog

Credit: Yanran Li, Hui Su, Xiaoyu Shen, Wenjie Li, Ziqiang Cao and Shuzi Niu,
*DailyDialog: A Manually Labelled Multi-turn Dialogue Dataset*, IJCNLP 2017.
[Paper](https://aclanthology.org/I17-1099/) and
[original dataset card](https://huggingface.co/datasets/li2017dailydialog/daily_dialog).
The source describes human-written everyday dialogues. This authorship designation does not independently
establish that every example records a spontaneous real-world interaction. Two speakers are assigned using the release's
alternating-turn convention, not verified real-world participant identities.

Downloaded source: [original-format mirror](https://huggingface.co/datasets/roskoN/dailydialog/tree/5214b2a66405abf87fd229e5c1007985501ffe3e),
revision `5214b2a66405abf87fd229e5c1007985501ffe3e`. The original website no longer serves its archive.
The three mirrored split ZIPs were verified against published sizes and SHA-256
hashes and parsed directly; the mirror's loader was not executed. Mirror
provenance is explicit: this is not an independently byte-verified copy fetched
from the defunct original host.

The original and mirror declare **CC BY-NC-SA 4.0**. This DailyDialog-derived
contribution is distributed under the same terms, with the transformation
notice below. [License](https://creativecommons.org/licenses/by-nc-sa/4.0/).
These source-specific terms do not assign a blanket license to this collection.

Changes: first retain only the first block in eight source records that repeat
the entire dialogue (seven train, one validation); 67 repeated turns removed.
The retained first-block wording and annotations remain unchanged. Then create
two perspectives per retained dialogue; each target human's speech
becomes `assistant` and the other speaker becomes `user`. Removed `__eou__`
structural separators and their trailing non-message segment. All utterance
wording, punctuation, spelling, whitespace, boundaries and ordering are
preserved after the explicit repeated-block cleanup. No system messages, personas, speaker prefixes, dialogue-act labels
or emotion labels are inserted. Original annotations remain in local audits.
No speech or personas were generated.

Native input: 11,118 train, 1,000 validation and 1,000 test conversations.
Removed 746 complete normalized duplicate copies and 31 substantive complete-
prefix counterparts. Priority is test, then validation, then train; within a
split, keep the longest version, using original line order for ties. Complete
exact duplicate ties use the earliest source line. Every retained record stays
in its original split. Seven longer training copies were excluded to protect
shorter held-out conversations. Distinct continuations remain separate.

Retained conversations: 10,355 train; 991 validation; 995 test (12,341 total).
Exported perspectives: **20,710 train; 1,982 validation; 1,990 test; 24,682 total**.
Both perspectives of each conversation remain in the same split. IDs encode
original split, zero-based source line and target speaker (1 or 2).

Comparison uses NFC/whitespace normalization, merged same-speaker runs and
orientation canonicalization, ignoring system text only for comparison.
Prefixes must cover at least four turns and 200 normalized characters.
Retained records were checked against the existing published collection, with
no exact or qualifying full-prefix matches. This is not a guarantee against
semantic, interior-subsequence or participant-level leakage.

Every output message was compared with the source transformation; all three
splits were loaded using Hugging Face datasets. Five seeded conversations per
split, all eight repeated-block repairs, and all 31 prefix-removal decisions
were reviewed. Sampled semantic
review does not establish the quality of every utterance.

## Cornell Movie-Dialogs (`convokit/movie-corpus`)

Credit: Cristian Danescu-Niculescu-Mizil and Lillian Lee,
*Chameleons in Imagined Conversations: A New Approach to Understanding Coordination
of Linguistic Style in Dialogs* (CMCL at ACL 2011), and the ConvoKit distribution.
[Original source](https://www.cs.cornell.edu/~cristian/Cornell_Movie-Dialogs_Corpus.html)
and [ConvoKit documentation](https://convokit.cornell.edu/documentation/movie.html).
These are human-authored fictional script exchanges, not recordings of spontaneous
human conversations. One source conversation is a two-character reply chain; we
do not concatenate separate exchanges into a whole movie.

Input: the ConvoKit `movie-corpus.zip` archive, SHA-256
`c3581e79d31e55cc8242591e66f829e84c8bd98a2a9c017bd851382dcdc17ac2`,
40,854,701 bytes. The checksum was computed and pinned during inspection, not
advertised as an upstream published checksum. The download manifest advertises
version 2, while the archive index metadata says version 1; the content hash
identifies this input. The archive contains 83,097 conversations, 304,713
utterances, and 617 source movie IDs. Source files and metadata remain unchanged
locally. The original corpus README and source pages were inspected. **No explicit
corpus redistribution license was found** in that documentation. ConvoKit's
software license is not assigned to this data, and no blanket collection license
is assigned.

Conversion: each eligible individual character is a separate target perspective.
Its spoken text becomes `assistant`, with the other speaker as `user`. Collective
speakers remain context only. No system messages, personas, speaker prefixes,
movie metadata, ratings, or generated dialogue are inserted. Message boundaries,
ordering, consecutive same-speaker messages, original wording, meaningful
whitespace, literal spoken brackets, and genuine repetition are preserved except
for the documented conservative cleanup.

Cleanup removes verified `u`/`i`/`b` formatting tags, one trailing HTML footer,
blank messages, and a verified nonverbal belch marker; HTML entities are decoded
once. Reviewed literal angle-bracket speech is retained. Conversations containing
scene/control/narrator identities are excluded. Damaged or ambiguous extracted
text and unresolved identities stay outside the published data. Each retained
conversation still contains both speaking participants and at least two spoken
messages. All exceptions are bound to audited source IDs and the pinned archive.

Existing SPOLIN data is unchanged. We exclude 18,675 incoming conversations with
verified SPOLIN pair overlaps, including pairs contained within longer exchanges.
We also exclude 329 internal complete duplicate copies, 83 scene/control/narrator
cases, 57 insufficient spoken exchanges, and one conversation without an
individual target. Another **146 conversations are quarantined** locally: 44
conflicting published speaker mappings, 43 uncertain identities, 43 damaged or
unresolved text cases, and 16 partial-overlap script variants. No qualifying
complete-prefix removals remained after those checks.

Comparisons ignore system text and perspective orientation, normalize Unicode
and whitespace, and recognize merged consecutive speaker runs without altering
exported text. Complete duplicates keep the lowest numeric movie/conversation ID.
Complete-prefix matching requires four merged runs and 200 normalized characters;
longer versions have priority, with IDs breaking ties. Interior matches of four
alternating runs and 200 characters are quarantined; known SPOLIN pairs are a
separate provenance-supported two-turn comparison. All 1,336,738 previously
published rows were checked. This does not establish a general semantic/fuzzy
or arbitrary-subsequence deduplication guarantee.

The source contains alternative or misattributed copies of some scripts. Four
verified source-ID pairs are grouped together: `m345→m58`, `m593→m225`,
`m275→m36`, and `m589→m588`. Raw titles, years, character IDs, and output IDs
are not rewritten. In particular, the `bloodmoon` source entry contains Blue
Velvet dialogue, and two U-Turn script versions have different source IDs.
The 617 IDs therefore form **613 groups**. Groups are ordered by SHA-256 of
`cornell-movie-split-v1:<canonical_movie_id>`; the first 30 form validation, the
next 30 test, and 553 train. Every perspective and conversation in a group stays
in one split. Because source groups have different sizes, row proportions differ
from nominal 90/5/5. Existing SPOLIN spans these films, so the combined collection
is **not movie-disjoint**. Source metadata may contain additional inaccuracies.

Retained: **63,806 conversations and 127,566 perspectives**:

| Split | Conversations | Perspectives |
| --- | ---: | ---: |
| Train | 57,792 | 115,556 |
| Validation | 3,019 | 6,027 |
| Test | 2,995 | 5,983 |

All 408,749 exported messages were compared with their prescribed source
transformations, including 204,383 target messages. All three splits load through
Hugging Face datasets. Twenty seeded conversations per final split and relevant
structural/markup edge cases were reviewed. Sampled semantic review cannot certify
every utterance; source typos, context-dependent fragments, and script-extraction
limitations remain. Detailed source hashes, exclusions, quarantine evidence,
lineage, grouping decisions, and review records stay local.


## EmpatheticDialogues

- Source: [Facebook Research EmpatheticDialogues](https://github.com/facebookresearch/EmpatheticDialogues), Hannah Rashkin, Eric Michael Smith, Margaret Li, and Y-Lan Boureau, *Towards Empathetic Open-domain Conversation Models: A New Benchmark and Dataset*, ACL 2019 ([paper](https://aclanthology.org/P19-1534/)).
- Terms: [Creative Commons Attribution-NonCommercial 4.0](https://github.com/facebookresearch/EmpatheticDialogues/blob/9649114c71e1af32189a3973b3598dc311297560/LICENSE). This source-specific license does not replace the other sources' terms or assign a blanket collection license.
- Original human-human dialogues imported directly from the authors' archive, not from OdysSim. Archive SHA-256: `56f234d77b7dd1f005fd365bb17769cfe346c3c84295b69bc069c8ccb83be03d`.
- Two perspectives per retained conversation. Target human speech is `assistant`; the other participant is `user`. Only the storyteller receives the original emotion and situation in `system`; listeners receive no system message. No generated personas.
- `_comma_` escapes decoded; literal quotes, wording, whitespace, chronological utterance boundaries, and native splits retained. Three single-message conversations excluded. In three individually approved training records, the identical second copy of a two-message exchange was removed. Ratings, tags, candidate responses, and other auxiliary fields are not rendered.
- 24,847 retained conversations yield 49,694 perspectives: 39,062 train, 5,538 validation, and 5,094 test. Both perspectives of each conversation remain together; original `valid` is published as `validation`.
- Native storyteller identities are separated between splits, but listener identities overlap. These two-perspective splits are not fully participant-disjoint. Exact/full-prefix overlap checking does not establish semantic or arbitrary-subsequence deduplication; manual semantic review is sampled.


## OpenAssistant

- Original sources: [OASST1](https://huggingface.co/datasets/OpenAssistant/oasst1/tree/fdf72ae0827c1cda404aff25b6603abec9e3399b) and [OASST2](https://huggingface.co/datasets/OpenAssistant/oasst2/tree/179dd21fc55192153d94adb0e0ce8f69e222bf75), by the OpenAssistant/LAION community; Andreas Köpf et al., *OpenAssistant Conversations – Democratizing Large Language Model Alignment* ([paper](https://arxiv.org/abs/2304.07327)). Original ready-for-export train/validation Parquets, not OdysSim conversions.
- Source-specific terms: Apache License 2.0, as declared by both original releases; [original license](https://huggingface.co/datasets/OpenAssistant/oasst1/blob/fdf72ae0827c1cda404aff25b6603abec9e3399b/LICENSE). No blanket collection license is assigned.
- English paths only. Each original human prompter becomes an assistant training target; original assistant replies become user context regardless of human/AI authorship. No system messages, generated personas, rewritten speech, custom masks, or special action tokens. Wording, whitespace, and native parent-chain order are preserved.
- Individual target provenance requires native synthetic=false and no contradictory model evidence. Deleted/rejected content, invalid paths and non-English messages are excluded; uncertain review/language status and contradictory message versions are quarantined locally. Metadata-based checks are not forensic proof of authorship or exhaustive linguistic verification.
- Releases jointly deduplicated, preferring OASST2 and retaining eligible OASST1-only content. Distinct human continuations survive; terminal assistant alternatives that add no new human target are reduced deterministically. Shared human prefixes intentionally repeat across retained branches; occurrence counts are kept locally, without weighting or masking.
- Entire equivalent tree groups remain together. Validation in either release takes precedence; a deterministic 5% of remaining eligible training groups forms test. Tree isolation does not imply author isolation: different volunteers may write successive prompter turns, and authors can occur across trees/splits. No stable personal identity is inferred.
- Source IDs, author IDs, branch decisions, provenance, exclusions, and audits remain local. Semantic review is sampled; exact and substantial complete-prefix comparisons do not establish semantic or arbitrary-subsequence deduplication.


### Humanual-News correction — 2026-11-12

Removed 38 previously published targets predominantly in Arabic script (Arabic/Persian): 36 train, 1 validation, and 1 test. Every other record and native split is unchanged. The exact IDs and target hashes are retained locally to prevent reintroduction. This targeted correction is not an exhaustive English-only certification of the remaining collection.


### Conservative language correction — 2026-11-15

Removed 121 previously published records from individually reviewed complete source paths containing clearly non-English human speech. All participant perspectives of those paths were removed together; retained text, ordering, roles and splits are unchanged. Detector-only flags, ambiguous code-switching, quoted material and foreign-context-only cases remain published. Removed records and exact IDs/hashes are retained locally for audit or future multilingual use. This is a limited correction, not an English-only certification. OpenAssistant's incorrect upstream language labels are documented in the local audit.

- BiTOD: 1 records removed ({'train': 1}).
- CaSiNo: 2 records removed ({'train': 2}).
- Humanual-Book: 12 records removed ({'train': 12}).
- Humanual-Chat: 8 records removed ({'train': 8}).
- Humanual-Email: 2 records removed ({'train': 2}).
- Humanual-News: 14 records removed ({'train': 6, 'test': 6, 'validation': 2}).
- Humanual-Politics: 7 records removed ({'train': 7}).
- LMSYS-Chat-1M: 17 records removed ({'train': 6, 'test': 5, 'validation': 6}).
- OpenAssistant: 21 records removed ({'validation': 3, 'train': 18}).
- TweetSumm: 8 records removed ({'test': 3, 'train': 5}).
- movie-corpus: 24 records removed ({'test': 4, 'train': 20}).
- spolin-corpus: 4 records removed ({'validation': 4}).
- wizard_of_internet: 1 records removed ({'train': 1}).