Multimodal College-level Writing Dataset
A dataset of authentic student essays, detailed instructor feedback, and transcripts of one-on-one student–instructor conferences. This is built to model how writing actually develops, not just how it's corrected.
Student essays
Multiple drafts of the same essay, collected as it's revised across a semester.
Instructor feedback
Detailed written comments instructors leave directly on each draft — an average of 11 per essay.
Conferencing dialogue
Transcripts of one-on-one conversations where the student and instructor discuss the essay and its feedback.
Intended impact
- Enable AI systems that model higher-order writing development, not just grammar.
- Provide a public benchmark for evaluating and training educational AI feedback models.
- Support equity research on how feedback effectiveness varies across ESL learners and demographics.
- Bridge learning sciences and NLP communities through open, multimodal data.
Possible Use Cases
- Automated essay scoringRubric-aligned, trait-specific assessment
- Feedback generationBenchmarking AI vs. authentic instructor feedback
- Revision outcome predictionPredicting draft improvement from feedback features
- Idea development modelingTracking central idea evolution across drafts
- Multimodal dialogue modelingTraining AI writing tutors on real conferencing data
- Feedback quality evaluationClassifying feedback by pedagogical function
Principal investigators
RC
Raquel Coelho (Co-PI)
Department of Informatics & Networked Systems / Learning Research & Development Center