Million Tutor Moves (MTM) Data Use Terms
Access is granted automatically when you accept the terms. We keep a record of everyone who has accepted them and will use the email on your Hugging Face account to tell you about corrections or withdrawn releases.
Million Tutor Moves (MTM) Data Use Terms
The Million Tutor Moves (MTM) data are published by the National Tutoring Observatory as a series of data releases under a Creative Commons Attribution 4.0 International (CC BY 4.0) license. Access is granted on the condition that you agree to the following terms. By requesting access you represent that you have read these terms, that you have authority to accept them on behalf of yourself and any organization you name, and that you will ensure they are observed by everyone to whom you provide the data. One acceptance covers every current and future MTM release published under these terms.
Important: these data have been de-identified, and de-identification is never perfect
Each release describes on its own page how it was de-identified. Some releases replace names, places, schools, dates and similar identifiers with labeled placeholders such as <NAME_1>; others replace them with realistic but fabricated substitutes. Two things are true in every case. First, no identifier in these data should be treated as a fact about a real person, place or institution. Second, because detection is imperfect, a small number of genuine identifiers may survive. The terms below govern both cases.
1. Attribution. You will cite the dataset in any publication, product documentation, or model card that draws on it, using the citation provided on each release's page.
2. No re-identification. You will not attempt to identify any student, teacher, tutor, classroom, school, or district represented in the data — directly, by inference, by linking these data with any other dataset or external information source, or by recovering any information that was removed or replaced — and you will not attempt to contact any individual represented in the data.
3. No recovering what was removed or replaced. You will not attempt, by any method, to determine what a placeholder or substitute stands for, or which identifiers in the data are genuine. This includes cross-referencing identifiers against external sources, rosters, public records, or web search; training or applying classifiers to separate genuine from replaced identifiers; and exploiting statistical or formatting artifacts of the de-identification process. These attempts are prohibited whether or not any individual is ultimately identified.
4. No harm to students, tutors or teachers. You will not use the data in any way that may harm students, tutors or teachers. This includes, but is not limited to: building models or systems that discriminate against students, tutors or teachers on the basis of race, ethnicity, gender, disability, language background, or any other demographic characteristic; and building systems that surveil, rate, discipline, or otherwise penalize tutors or teachers on the basis of their instructional practice.
5. No consequential decisions about individuals. You will not use the data, or any model or tool derived from it, to make or inform decisions about identifiable individuals, including decisions about employment, performance evaluation, discipline, admissions, placement, promotion, or eligibility for services.
6. Permitted uses, including commercial. Research, teaching, evaluation and benchmarking, product development, and the training or fine-tuning of models are all permitted, including by for-profit organizations, provided every other term here is met.
7. Protection against memorization and disclosure. If you train or fine-tune a model on these data and make that model available to others, you will take reasonable, documented measures to prevent it from reproducing verbatim passages of the transcripts, and you will test for this before release. De-identification does not remove this obligation, since verbatim reproduction may still expose residual genuine identifiers.
8. Reporting suspected genuine identifiers. If you come to believe that a particular identifier in the data is genuine — for example because you recognize an individual, or because you encounter it incidentally in the course of ordinary use — you will notify us promptly at sandpiper_admin@cornell.edu, will not disclose it to anyone else or publish it, and will delete or redact the affected material on request. This obligation applies to incidental discovery only; it does not authorize the searching described in Section 3.
9. Secure handling. You will store the data securely and will ensure that any collaborator, employee, contractor, or student to whom you provide access is bound by these same terms.
10. Publication of excerpts. If you quote transcript excerpts in publications, documentation, or demonstrations, you will reproduce the de-identified text exactly as it appears in the data, including any placeholders or substitutes, and will not substitute, verify, or comment on the authenticity of any identifier.
11. Redistribution. CC BY 4.0 permits redistribution. We ask that you link to the release page rather than re-hosting the files, so that others receive the current version, these terms, and any subsequent corrections. If you do share a copy, we ask that you include the release name and version, a link to its page, and any correction notice we have published for it.
12. Sensitive content. These are authentic instructional interactions and may contain content that is sensitive or distressing. Consider this in the design of any human-facing study or product built on them.
13. Your own compliance obligations. You are responsible for compliance with applicable law, your institution's policies, and any human-subjects review requirements that apply to your use.
14. No warranty. The data are provided "as is," without warranty of any kind, express or implied. No representation is made that de-identification is complete.
15. Termination. We may revoke access at any time for violation of these terms, and may require that you delete your copies of the data. Revocation does not relieve you of obligations under Sections 2, 3, 4, 5, 7, and 8. Revocation applies to every MTM release.
16. Corrections and withdrawn releases. If we withdraw a release, or part of one, or publish a corrected version, we will notify everyone who has accepted these terms at the email address on their Hugging Face account. On notice you will stop using the withdrawn material, delete every copy you hold, including derived files and backups, within 10 days, and confirm the deletion in writing if we ask. You may continue to use corrected versions.
17. Changes to these terms. We may revise these terms for future releases. The version you accepted continues to govern the releases you accessed under it. If a revision changes your obligations, we will ask you to accept it again before you access further releases.
18. Combining data and sharing results. You may combine MTM releases with one another. Sections 2 to 5 continue to apply to any combination of these data with other data, and to any query results, aggregates, models, or derived datasets you share.
Log in or Sign Up to review the conditions and access this dataset content.
Million Tutor Moves, release 0 (sample)
Release: mtm-000-sample version 1.0 · Published: 2026-11-16 · Contact: sandpiper_admin@cornell.edu
Release 0 is a preview. It contains 4 tutoring sessions from different programs, so researchers can see the format of the Million Tutor Moves release before the full data ships. Every session was de-identified by hand (see De-identification).
Million Tutor Moves is a collection of tutoring interactions assembled by the National Tutoring
Observatory (NTO) at Cornell University. Each session is described by tutoring context
variables on the sessions table:
| column (JSON) | values | meaning |
|---|---|---|
context_tutor_type (context.tutor_type) |
human, ai |
who tutors: a human tutor or an AI tutor |
context_group_format (context.group_format) |
one_on_one, group |
one tutor with one student, or with a small group |
context_delivery_model (context.delivery_model) |
high_dosage, on_demand, assignment_specific |
how and when the student gets tutoring: scheduled, recurring program sessions; help the student asks for when they need it; or tutoring tied to a specific assignment |
Filter on any combination, for example human one-on-one high-dosage tutoring, or all on-demand
help. More context variables may be added in later releases, always with the context_ prefix.
Contents
context_tutor_type |
context_group_format |
context_delivery_model |
sessions | segments | events |
|---|---|---|---|---|---|
human |
group |
high_dosage |
1 | 1,083 | 20 |
human |
one_on_one |
high_dosage |
1 | 317 | 0 |
human |
one_on_one |
on_demand |
2 | 44 | 10 |
- A segment is one spoken or typed turn.
- An event is a non-speech item: a whiteboard action, an AI-generated description of the shared screen, or a quiz answer.
Schema structure
Each session is stored as three linked tables:
sessions (one row per session)
session_id ──┬──< segments (one row per spoken or typed turn)
│ session_id, segment_index, role, speaker_id, content, annotation_* ...
└──< events (one row per non-speech item)
session_id, event_index, type, role, speaker_id, content, meta, ...
sessionsdescribes the session as a whole: its context variables, subject and level, how it was conducted, its length and counts, who took part (tutor_id,student_ids), its outcomes and its annotations.segmentsis the dialogue: one row per turn, in order (segment_index), with who spoke (role,speaker_type,speaker_id), when (started_at,ended_at) and what was said (content), plus any annotations of that turn.eventsholds everything that isn't dialogue, on the same timeline: whiteboard actions, an AI-generated description of the shared screen, or a quiz answer.typesays which, andmetaholds details that depend on the type.
Keys and joins:
session_idis the primary key ofsessions.(session_id, segment_index)is the key ofsegments, and(session_id, event_index)is the key ofevents.- A participant's
speaker_idis the same insegmentsandevents.sessions.tutor_idandsessions.student_idslist thespeaker_ids of the session's tutor and students. - Context variables, outcomes and session annotations are on
sessionsonly. To filter segments or events by them, join tosessionsonsession_id(see Loading).
Variable groups: related variables are grouped so new ones can be added without reshaping the tables. In Parquet and CSV a group is a set of flat columns sharing a prefix; in JSON it is a nested object.
| table | group | Parquet / CSV columns | JSON | in release 0 |
|---|---|---|---|---|
| sessions | context | context_* |
context: {...} |
tutor_type, group_format, delivery_model |
| sessions | outcomes | outcome_* |
outcomes: {...} |
next_problem_correct |
| sessions | annotations | annotation_* |
annotations: {...} |
none yet (empty object, no columns) |
| segments | annotations | annotation_* |
annotations: {...} on each segment |
none yet (empty object, no columns) |
Annotations are labels added to the data after the session, by human annotators or models, for example codes for tutoring moves. Session annotations describe the whole session; segment annotations describe one turn. Release 0 has none, so the groups exist only as empty objects in JSON. Later releases add variables to these groups (for example, annotations or more outcome measures) without renaming existing ones.
Conventions:
- Time:
started_at/ended_atare seconds from the start of the session. There are no dates or clock times. Quiz answers can have negative times when they happened before the chat. - Speakers:
roleis a label (Tutor,Student,Student 2), never a name.speaker_typeistutor,student,ai_tutororother. - IDs:
session_id,speaker_id,tutor_idandstudent_idsare generated by NTO and can't be linked back to the source systems. Every tutor and student has a person ID. In release 0, person IDs are scoped to one session, so they don't link a person across sessions. - Speakers without a person ID:
speaker_idis null for AI tutors and forotherspeakers:Background(third parties, noise) andUnknown(a speaker who couldn't be identified, such as a student the transcription couldn't match to a numbered student). There's deliberately no placeholder ID such asp_unknown, because it would be counted as a person when grouping or counting byspeaker_id. To find these turns, filter onroleorspeaker_type. - Missing values are null (empty in CSV). A null means the source doesn't record that field for the session; it doesn't mean zero or false.
Files
| format | path | layout |
|---|---|---|
| Parquet | data/parquet/sessions.parquet, segments.parquet, events.parquet |
one file per table |
| CSV | data/csv/sessions.csv, segments.csv, events.csv |
same as Parquet; list and object columns are JSON strings |
| JSON Lines | data/json/sessions.jsonl |
one session per line, with its segments and events nested as lists |
The schema files call this release v0. The formal definition of one session is the JSON
Schema in schema/mtm_session.schema.json; this card's data
dictionary is also in schema/fields.md.
Loading
There are three tables, each a config: sessions (the default), segments and events.
from datasets import load_dataset
repo = "NationalTutoringObservatory/mtm-000-sample"
sessions = load_dataset(repo, "sessions", split="train")
segments = load_dataset(repo, "segments", split="train")
# Segments from human, group, high-dosage sessions
keep = set(sessions.filter(lambda r: r["context_tutor_type"] == "human"
and r["context_group_format"] == "group"
and r["context_delivery_model"] == "high_dosage")["session_id"])
group = segments.filter(lambda r: r["session_id"] in keep)
Run hf auth login first: access requires accepting the Data Use Terms.
-- DuckDB, or the SQL Console on this page
SELECT s.context_tutor_type, s.context_group_format, s.context_delivery_model, g.role, g.content
FROM 'hf://datasets/NationalTutoringObservatory/mtm-000-sample/data/parquet/segments.parquet' g
JOIN 'hf://datasets/NationalTutoringObservatory/mtm-000-sample/data/parquet/sessions.parquet' s
USING (session_id)
WHERE s.context_delivery_model = 'on_demand'
ORDER BY g.session_id, g.segment_index;
Data dictionary
Every field in every table. Types: "or null" means the field can be empty.
sessions
One row per session. Groups: context (context_), outcomes (outcome_) and annotations (annotation_). Grouped variables are flat columns with the group's prefix in Parquet and CSV, and nested objects in JSON; the "in JSON" column gives each one's path.
| column | in JSON | type | description |
|---|---|---|---|
session_id |
session_id |
string (session ID) | Session ID. Primary key of sessions; joins segments and events to their session. |
context_tutor_type |
context.tutor_type |
string: human, ai |
Who tutors: a human tutor, or an AI tutor. |
context_group_format |
context.group_format |
string: one_on_one, group |
One tutor with one student (one_on_one), or with a small group of students (group). |
context_delivery_model |
context.delivery_model |
string: high_dosage, on_demand, assignment_specific |
How and when the student gets tutoring. high_dosage: scheduled, recurring sessions as part of a tutoring program. on_demand: help the student asks for when they need it. assignment_specific: tutoring tied to working through a specific assignment. |
subject |
subject |
string or null | Broad subject, e.g. math, chemistry. Null when unknown. |
topic |
topic |
string or null | Narrower topic within the subject, e.g. a unit or course name. Null when unknown. |
knowledge_components |
knowledge_components |
list of string | Skills or concepts the session targets (question tags, construct names or learning objectives, depending on the source). Empty when unknown. |
school_level |
school_level |
string: primary, lower_secondary, upper_secondary, postsecondary or null |
Student's school level on the UNESCO ISCED 2011 scale: primary (1), lower_secondary (2), upper_secondary (3), postsecondary (5+). Null when unknown. |
modality |
modality |
string: video, voice, text_chat, mixed or null |
How the session was conducted: video, voice, text_chat or mixed. Null when unknown. |
n_students |
n_students |
integer | Number of students taking part. |
duration_s |
duration_s |
number or null | Session length in seconds: the latest segment time (ended_at, else started_at). Null when the session has no timing. |
n_segments |
n_segments |
integer | Number of rows in segments for this session. |
n_events |
n_events |
integer | Number of rows in events for this session. |
n_speakers |
n_speakers |
integer | Number of distinct segment roles (speaker labels). |
tutor_talk_share |
tutor_talk_share |
number or null | Share of segment words spoken or typed by tutors (speaker_type tutor or ai_tutor), 0-1. Null if there are no words. |
tutor_id |
tutor_id |
string (person ID) or null | Person ID of the session's tutor; equal to speaker_id on the tutor's rows. Null when there is no human tutor. |
student_ids |
student_ids |
list of string (person ID) | Person IDs of the session's students; the distinct speaker_id values on student rows. |
outcome_next_problem_correct |
outcomes.next_problem_correct |
boolean or null | Whether the student's first answer after the session, to a new question (not a retry) on the same skill as the question answered just before it, was correct. Null when unavailable. |
none yet (annotation_*) |
annotations (empty object) |
Annotations: labels added to the data after the session (for example, codes for tutoring moves), by human annotators or models. Session annotations describe the whole session; segment annotations describe one turn. Empty in this release; new annotations are added to this group. |
segments
One spoken or typed turn. One row of the segments table. Groups: annotations (annotation_). Grouped variables are flat columns with the group's prefix in Parquet and CSV, and nested objects in JSON; the "in JSON" column gives each one's path.
| column | in JSON | type | description |
|---|---|---|---|
session_id |
session_id |
string (session ID) | Session the segment belongs to. |
segment_index |
segment_index |
integer | Position of the segment in the session, from 0. Segments are ordered by this. |
started_at |
started_at |
number or null | When the turn starts, in seconds from the start of the session. Null when the source has no timing. |
ended_at |
ended_at |
number or null | When the turn ends, in seconds from the start of the session. Null when the source has no end times. |
role |
role |
string | Speaker label within the session, never a name: Tutor, Student, Student 2, AI, Background or Unknown. 'Unknown' is a speaker that couldn't be identified: an unusable source label, or in a group session a student the transcription couldn't match to a numbered student. |
speaker_type |
speaker_type |
string (speaker type) | Category of the speaker (see speaker_type). |
speaker_id |
speaker_id |
string (person ID) or null | Person ID of the speaker. Null for AI tutors and for 'other' speakers (Background, Unknown): there is no placeholder ID for unidentified speakers, so they are never counted as a person. |
content |
content |
string | What was said or typed, de-identified. Speech transcripts may contain markers such as [unclear]. |
none yet (annotation_*) |
annotations (empty object) |
Annotations: labels added to the data after the session (for example, codes for tutoring moves), by human annotators or models. Session annotations describe the whole session; segment annotations describe one turn. Empty in this release; new annotations are added to this group. |
events
One non-speech item: a whiteboard action, an AI-generated description of the shared screen, or a quiz answer. One row of the events table.
| field | type | description |
|---|---|---|
session_id |
string (session ID) | Session the event belongs to. |
event_index |
integer | Position of the event in the session, from 0. |
started_at |
number | When the event happens, in seconds from the start of the session. May be negative for items before the first segment (e.g. quiz answers before the chat). |
ended_at |
number | When the event ends, in seconds from the start of the session (equal to started_at for instantaneous events). |
role |
string or null | Label of the participant who caused the event, as in segments.role. Null when no participant did (e.g. visual descriptions). |
speaker_type |
string (speaker type) or null | Category of the participant who caused the event. Null when role is null. |
speaker_id |
string (person ID) or null | Person ID of the participant who caused the event, matching their segments.speaker_id. Null when role is null. |
type |
string | Kind of event: a whiteboard action name (e.g. AddCanvasObjectEvent), visual_description or quiz_answer. |
content |
string | Description of the event: the whiteboard action, the AI-generated screen description, or the quiz question's skill. |
text |
string or null | Text a participant wrote on the whiteboard (sticky notes, text boxes), when the event has any. Null otherwise. |
meta |
object | Other details that depend on the event type (see the meta keys table). JSON object; a JSON-encoded string in CSV. |
meta keys by event type
event type |
meta key |
description |
|---|---|---|
AddCanvasObjectEvent |
object_type |
Kind of whiteboard object acted on, e.g. STICKER, TEXT, SHAPE. |
quiz_answer |
answer_context |
Part of the quiz the answer was given in: Checkin, Checkout, or a retry of either (CheckinRetry, CheckoutRetry). |
quiz_answer |
is_correct |
Whether the answer was correct (true/false). |
visual_description |
subject |
What the description is of: Whiteboard or Student. |
Shared types
- session ID: s_ followed by 16 hex characters. Generated by NTO; can't be linked back to source systems.
- person ID: p_ followed by 16 hex characters. Generated by NTO; can't be linked back to source systems. In release 0, IDs are scoped to one session: the same person in two sessions has two IDs.
- speaker type: Speaker category: tutor, student, ai_tutor, or other (a speaker who is neither tutor nor student, or couldn't be identified: 'Background' noise or an 'Unknown' speaker).
De-identification
Each session's text was de-identified by hand. Personal information was replaced with realistic surrogates rather than tags:
- Names, places and organizations were replaced with plausible substitutes. A name you see in the text is not the person's real name.
- Platform and organization names were replaced too, with fictional names (for example, "Lumora").
- Some incidental details were altered (for example, a grade level mentioned in conversation). Don't treat conversational details as facts about the participants.
- Speaker labels and person IDs carry no identity.
NTO then reviewed every row of release 0 for remaining personal information. That review made a few further corrections, such as giving every mention of a person the same surrogate and altering an age.
Before release, NTO also checked the text for e-mail addresses, web addresses and phone numbers. None remain, apart from two web addresses that belong to the fictional platform name. One other match was checked by hand and is a false positive (a list of prices that looks like a phone number).
De-identification is never perfect. If you find an identifier that looks genuine, see section 8 of the Data Use Terms.
Known limitations
- Release 0 is 4 sessions. It shows the format, not the full data, and isn't suitable for
statistical analysis. There are no AI-tutored (
context_tutor_type = ai) sessions in release 0 yet. - Speech transcripts and speaker labels are automatic and imperfect. In spoken sessions, the
text comes from speech recognition, and who-said-what (
role,speaker_type,speaker_id) comes from speaker diarization. Its accuracy depends on the quality of the audio and on the diarization models used. Expect some turns to be attributed to the wrong speaker, one speaker to be split into two labels, or two speakers merged into one, especially with overlapping speech, background noise or several students on one microphone. Treat speaker-level measures (such astutor_talk_shareor per-student counts) as estimates. - Human, group, high-dosage session: the transcript comes from automatic speech recognition, with speakers
labeled by diarization. Turns diarization couldn't match to one of the numbered students
are labeled
Unknown(speaker_typeother) and have no person ID.visual_descriptionevents were generated by an AI model from screenshots of the shared whiteboard. - Human, one-on-one, high-dosage session: the transcript comes from speech recognition and contains markers
such as
[unclear]. In 9 turns the recognizer switched speaker partway through; those speaker labels were removed during review, so these turns may hold more than one speaker's words. Timing is to the second, and there's noended_at. - On-demand sessions:
- One is a text chat with no timing, topic or grade level in release 0.
- The other is one tutor chat inside a quiz attempt, and that attempt's quiz answers are included as events. Some answers came days after the chat.
Versions
| Version | Date | Change |
|---|---|---|
| 1.0 | 2026-11-16 | First publication. |
Citation
@misc{mtm_release0_2026,
title = {Million Tutor Moves, release 0 (sample)},
author = {{National Tutoring Observatory}},
year = {2026},
publisher = {Hugging Face},
url = {https://huggingface.co/datasets/NationalTutoringObservatory/mtm-000-sample}
}
License and terms
Data: CC BY 4.0. Please credit the National Tutoring Observatory and cite the release as above. The data partners aren't named in release 0.
Access is granted on acceptance of the Million Tutor Moves Data Use Terms shown above (also in DATA_USE_TERMS.md). Please link to this page rather than re-hosting the files, so others get the current version and any corrections.
Release record
| Item | Value |
|---|---|
| Source package | NTO v0 package, built 2026-11-05; revised 2026-11-11 (text column typed, annotation groups added) and 2026-11-12 (a name and speaker markers removed from one session) |
| De-identification | by hand, with surrogates; then an NTO review of every row, completed 2026-11-04 |
| File hashes | see SHA256SUMS |
| Sign-off | Kirk Vanacore (data) and Doug Pietrzak (publishing), 2026-11-16 |
- Downloads last month
- 20