Question-Anchored-Tutoring-Dialogues-2k

This dataset contains dialogues from math tutoring interventions recorded on Eedi.

Dataset Details

Dataset Description

Each dialogue represents a chat-based conversation between a tutor and a student prompted by the student requesting assistance while working on a lesson. Dialogues are accompanied with 2 sources of meta-data:

1. DQ-Question-Metadata: The question the student was working on that prompted the tutoring session.
2. Dialogue-Subjects: The subject, topic(s), and subtopic(s) of the lesson the student was working on.

• Curated by: Matthew Zent, Digory Smith, and Simon Woodhead
• Funded by: Eedi
• Language(s) (NLP): English
• License: cc-by-nc-sa 4.0

Dataset Sources [optional]

• Repository: Eedi/Question-Anchored-Tutoring-Dialogues-2k
• Paper [IN PROGRESS]
• Demo: Eedi/question-anchored-tutoring-dialogues

Uses

Direct Use

This dataset is intended for non-commercial research focused on improving learning outcomes. Suitable use cases include, but are not limited to, model training, fine-tuning, and calibration tasks focused on effective dialogue modeling and tutoring interactions in educational contexts.

Researchers are encouraged to explore how dialogue-based interventions support student understanding and to use this data in ways that positively contribute to learning outcomes.

For commercial use, please contact us at hello@eedi.com.

Out-of-Scope Use

Any attempts to identify or re-identify individuals represented in the data are not allowed.

Additionally, use cases that risk adverse outcomes for students or educators, including but not limited to surveillance, profiling, or automated high-stakes decision-making without human oversight, are considered out of scope.

Dataset users are expected to carefully consider the impact of their work and avoid applications that could reinforce bias or inequality in learning environments.

Dataset Structure

Eedi/Question-Anchored-Tutoring-Dialogues-2k/
├── README.md
├── anchored-dialogues/
│   ├── train.csv
│   ├── test.csv
├── dialogue-subjects.csv
├── dq-question-metadata.csv
└── dataset_config.yaml

anchored-dialogues

Each dialogue represents a conversation between a tutor and stutent on Eedi. The conversation started when the student asked for help while working on a Diagnostic Question.

Glossary

Term	Definition
InterventionId	A unique identifier for the intervention.
TutorId	Unique identifier for a tutor.
QuestionId_DQ	The QuestionId of the Diagnostic Question (DQ) that the intervention is anchored to.
MessageSequence	The sequence number of the message in the entire dialogue.
IsTutor	Defines if a message is from a tutor (1) or a student (0).
MessageString	The message content.
TalkMovePrediction	GPT-applied labels for prompts by the tutor used to support students’ mathematical thinking, understanding, and communication (see Talk Move Annotations).

dq-question-metadata

Each Diagnostic Question (DQ) is presented to students as a single image, with the question and answer options embedded within the image. We have extracted the text associated with the question and answer options. Where the question or answer contained an image, the extracted text is a description of the image.

Glossary

Term	Definition
QuestionId_DQ	The QuestionId of the Diagnostic Question (DQ) that the intervention is anchored to.
InterventionId	A unique identifier for the intervention this DQ is anchored to.
MetaDataId	Unique metadata identifier.
Text	Extracted text for the question. LaTeX is denoted as \(...\). The text for image-based items is a description of the image.
Sequence	The order in which the text is presented.
MetaDataTagId	An integer value with a 1:1 mapping to the 'Label' column.
Label	A label denoting what the extracted text refers to. e.g. Answer B Text.

dialogue-subjects

The subjects/topics of tutor-student dialogues.

The subjects are organised in a hierarchy, with increasing granularity as the SubjectLevel increases. The ParentSubjectId allows us to navigate the hierarchy.

Dialogues without a subject can occur when a student has asked for help outside of a lesson. For example when asking for help with non-Eedi homework.

Glossary

Term	Definition
InterventionId	A unique identifier for the intervention.
SubjectId	Unique subject identifier.
ParentSubjectId	Unique parent subject identifier.
SubjectName	The name of the subject.
SubjectLevel	The level of the subject.
SubjectType	The type of the subject, one of 'Subject', 'Topic', or 'Subtopic'.

Dataset Creation

Source Data

Dialogues and all learning content are sourced from the math learning platform Eedi. Anchored-Dialogues include a column 'TalkMovePrediction' used to facilitate [Data Processing](#data collection and processing). Talk move predictions were sourced from Moreau-Pernet et al. (2024) (see [Talk Move Annotation](#talk moves))

Data Collection and Processing

Conversation data and associated metadata were sourced from Eedi's internal databases from Nov. 2021 to Feb. 2025.

The dataset was curated through the following sequential filtering and processing steps:

1. Initial Filtering
   Dialogues were selected if they contained at least 20 total messages, with a minimum of 7 messages from both the student and the tutor.

2. Content Moderation
   Conversations were screened using OpenAI’s omni-moderation-latest model across the following safety categories:
   "sexual", "sexual/minors", "harassment", "harassment/threatening", "hate", "hate/threatening", "illicit", "illicit/violent", "self-harm", "self-harm/intent", "self-harm/instructions", "violence", "violence/graphic".

3. Tutor Consent
   Only dialogues involving one of 25 tutors who provided explicit consent for research use were retained.

4. Downsampling
   The dataset was downsampled to 2,000 conversations, with a cap of 1,000 unique questions and no more than 8 conversations per question. Sampling prioritized diversity in dialogue quality and structure by maximizing the TF-IDF score of each conversation's TalkMovePredictions.

5. Manual Review
   Conversations and question texts were manually reviewed for potential personally identifiable information (PII) and content appropriateness. A total of 29 conversations were removed due to safety concerns or lack of educational value.

6. Potential-PII Anonymization Potential-PII was anonymized using PIIvot, which utilizes a hidden-in-plain-sight approach to replace labeled text with surrogates that maintain the context of the dialogue.

   Names, date of births, emails or social handles, urls, locations or addresses, phone numbers, and school names are replaced with realistic surrogates (see Potential-PII Annotations for how labels were applied).

7. Train/Test Splits Train/test splits are obtained using a 0.8/0.2 split on the dialogue key 'InterventionId'.

Who are the source data producers?

This dataset originates from conversations between Eedi tutors and students to support student learning while completing online math lessons. Student demographic information is reported by student whereas platform informormation is by-intervention at the time of each intervention and includes multiple rows per student.

Overview

Statistic	Count
Unique students	1,073
Unique tutors	25
Unique Interventions	1,971
Total messages	68,717

Student Demographics

Category	Distribution
Gender	Female: 314, Male: 159, Unknown: 600
Income Status	Low-income: 173, Not low-income: 480, Unknown: 420
Location	UK: 856, Unknown: 217

Platform Information

Category	Average (25%, 75%)
DaysOnPlatform	140.5 (9, 210.3)
HistoricQuestionCount	244 (25, 228)
HistoricCorrectness	0.62 (0.54, 0.72)
AgeAtIntervention	12 (11, 13)

Annotations

Talk Moves

To facilitate downsampling, talk moves were applied using the GPT-based classification model in Moreau-Pernet et al. (2024). The model was fine-tuned on conversation transcripts from small-group math tutoring sessions.

To evaluate the model’s generalizability to 1:1 chat-based math tutoring sessions, the 1st author manually annotated a validation set of 200 tutor utterances sampled through weighted stratified sampling using the orignal codebook from Moreau-Pernet et al. (2024). The sample distribution was flattened by 0.8 of the original label distribution represented in the full 4k dialogue dataset in order to validate more examples of minority class labels.

An error analysis was conducted on all mismatched labels. All instances where the model predicted a <Getting Student to Relate> label involved a common type of Eedi's word problems involving a conversation between two fictional students. The original codebook was not designed with this unique edge-case in mind and as such we posit the <Getting Student to Relate> label may not generalize this to 1:1 context. Considering this limitation, we present an additional results excluding GSR labels and find similar acceptable metrics to the original paper.

Label	Ratio of 4k Dialogues	Ratio of Validation Set	Baptiste et al.	F1 Score (Original)	F1 Score Excl. <GSR>	F1 Score Baptiste et al.
<None>	0.42	0.28	0.73	0.8991	0.8991	0.96
<Keep Together>	0.36	0.20	0.09	0.8333	0.8750	0.81
<Revoicing>	0.12	0.18	0.03	0.8986	0.8986	0.76
<Press for Accuracy>	0.06	0.16	0.13	0.7733	0.8286	0.88
<Getting Student to Relate>	0.02	0.08	0.004	0.0000	--	0.75
<Press for Reasoning>	0.002	0.06	0.006	0.7857	0.9565	0.94
<Restating>	0.0003	0.04	0.008	0.8000	0.8000	0.95

Potential-PII

To facilitate PII anonymization, both machine and human-annotated labels were applied to the data.

A labeling codebook was developed to label an independent set of ~40k student-tutor messages, achieving a minimum macro-averaged F1 score of 0.98 between raters over ~350 dialogues. 4 annotators from this original labeling initiative then applied the codebook to QATD-2k.

Machine labels for potential-PII were obtained using the PIIvot package and a DeBERTa-based NER pipeline fine-tuned on the prior set of ~40k student-tutor messages. The NER model was trained to recognize seven types of entities that may contain Personally Identifiable Information (PII): names, date of births, emails or socials, urls, locations or addresses, phone numbers, and school names.

Disagreements between machine and annotator labels were resolved to determine ground truth labels. Micro-averaged metrics for both PIIvot and Annotator label sets in relation to this ground truth are included below.

Label Set	Precision	Recall	F1 Score
Dialogues
PIIvot	0.984	0.984	0.984
Annotators	0.993	0.995	0.994
Questions
PIIvot	0.991	0.699	0.820
Annotators	0.997	0.997	0.997

We make three important distinctions:

1. Potential-PII: Our approach does not attempt to differentiate between real and fictional entities when labeling for potential PII. This simplifies the NER objective by avoiding the need to resolve whether a given name is "real." We instead rely on the anonymization step to enforce dialogue coherence (i.e., ensuring that a name anonymized in the question remains consistent throughout the dialogue). While this is suitable for Eedi's math word problems, it may be less appropriate in domains where specific names carry meaning (e.g., a historical essay).

2. Performance degradation on LaTeX questions: The lower performance of the model on question metadata (compared to dialogues) is attributed to domain shift. The NER model was trained on dialogue data and struggles to generalize to LaTeX formated question text.

3. NER Model release constraints: Due to the sensitive nature of the labeling task, the fine-tuned NER model used in this process cannot be released publicly due to re-identification concerns. It could potentially be used to detect residual PII in datasets processed using the HIPS anonymization method.

Personal and Sensitive Information

This dataset contains real student-tutor dialogues. Extensive efforts were made to mitigate privacy risks, but these risks cannot be alleviated completely. Identifying details were removed or anonymized where detected, but some sensitive content may persist. If you have concerns about any content, please contact the first author at matthew.zent@eedi.co.uk.

Bias, Risks, and Limitations

We present aggregate demographic and behavioral statistics to help downstream users assess potential biases in the dataset. Here we note highlight high-impact biases and limitations for downstream use:

1. QATD’s focuses on UK students and intentionally excludes US-based learners.
2. The data reflects real interactions on the Eedi platform, where tutors sometimes manage multiple students simultaneously, especially during peak hours.
3. The collection period includes growing public exposure to AI chatbots. As such, there are many instances of students expressing skepticism about whether their tutor was a real person.
4. Our strong commitment to student privacy limits access to individual-level student details, which may constrain downstream use in student modeling tasks.

Citation [optional]

Publication in Progress. Please reach out to Matthew.Zent@eedi.co.uk for citation information.

Dataset Card Author

Matthew Zent

Dataset Card Contact

Matthew.Zent@eedi.co.uk