Public Goods
The Inventory of Public Goods is a public repository of the resources that the EDSI team has created. These resources include data sets, code, frameworks, whitepapers, journal articles, and more. This inventory will be updated frequently. Resources are grouped by type.
Sign up for our newsletter to get updates on new resources!
Pairadigm
Pairadigm is an open-source Python library for producing continuous measurement scales from text using large language models.Pairadigm is an open-source Python library for producing continuous measurement scales from text using large language models. It implements a Concept-Guided Chain-of-Thought (CGCoT) methodology that surfaces nuance in each text and generates reasoned pairwise comparisons, judging which of two items expresses more of a target construct, rather than relying on direct scalar ratings. Supporting OpenAI, Anthropic, Google Gemini, and local models via Ollama and Hugging Face, it lets researchers validate LLM annotations against a small set of human labels (using diagnostics such as the Alternative Annotator Test) and, once validated, scale up to annotate far larger datasets affordably. The comparisons are modeled with a Bradley-Terry framework to recover continuous scores, and reward models can be fine-tuned to extend measurement to new data. Pairadigm is installable via pip install pairadigm
Find the code on GitHub: https://github.com/mlchrzan/pairadigm
Student Idea Attribution (SIA)
Student Idea Attribution occurs when a teacher takes up a student’s mathematical idea by revoicing, restating, or calling attention to it and explicitly attributes it to a specific student.
Mathematical Vocabulary
A combined glossary of mathematical vocabulary terms.
EmoPair
The EmoPair models are released as an open collection on Hugging Face, comprising two RoBERTa-large emotion-scoring models fine-tuned on the EmoPair annotations. PERT-EmoPair uses a Prompting-based Emotion Regression architecture, while Reward-EmoPair uses a reward-based architecture trained on the dataset's pairwise comparisons. Both predict continuous Valence, Arousal, and Dominance scores for English text and outperform equivalent models trained on the original EmoBank averages, particularly on the harder Arousal and Dominance dimensions.
The models are available at: https://huggingface.co/collections/edsi-umd/emopair.
Github: https://github.com/EDSI-UMD-College-Park/EmoPair
RealClass: A Framework for Classroom Speech Simulation with Public Datasets and Game Engines
RealClass tackles a major bottleneck in educational technology: AI speech models struggle to understand teachers and students in noisy, echoey classrooms, but gathering real audio is severely restricted by vital privacy laws. To solve this data shortage, this paper introduces an open-source framework that uses 3D game engines to realistically simulate classroom acoustics and background noise. By pairing these synthetic environments with public audio of children and instructional speech, the resulting RealClass dataset allows developers to train highly accurate classroom AI models safely and equitably, providing a powerful, privacy-first public good for the future of education.
MPT
M-Powering Teachers provides a dynamic platform dedicated to enhancing the professional growth and effectiveness of teachers. Through machine learning-based automated feedback on mathematics lessons, personalized support, and continuous improvement research, we aim to empower educators to excel in their practice, inspire lifelong learning, and positively impact student success. We strive to cultivate a thriving ecosystem where teachers can access the tools, knowledge, and connections needed to help each student reach their full potential.
Introduction to the tool: https://edsi.umd.edu/projects/m-powering-teachers/how-it-works
Explore the tool here: https://mpoweringteachers.stanford.edu/auth/login?callbackUrl=%2F
EmoPair
EmoPair improves the emotional dimension annotations in EmoBank, a widely used corpus of roughly 10,000 English sentences labeled along the Valence-Arousal-Dominance framework. EmoPair replaces direct scalar ratings with a scalable, LLM-driven pairwise comparison approach, in which annotators and models judge which of two sentences expresses more of a given dimension. Using Concept-Guided Chain-of-Thought (CGCoT) prompting, automated judgments are generated at scale, validated for human-level replacement via the Alternative Annotator Test (AltTest), and converted into continuous scores with a Bradley-Terry model. Models trained on EmoPair outperformed EmoBank-trained baselines on the harder Arousal and Dominance dimensions.
Find the GitHub repo introducing EmoPair and the dataset here: https://github.com/EDSI-UMD-College-Park/EmoPair
NCTE
The NCTE Transcripts Dataset addresses the scarcity of large-scale instructional corpora by providing the largest public dataset of mathematics classroom transcripts available to researchers. Compiled by the National Center for Teacher Effectiveness, it contains 1,660 anonymized, 45-to-60-minute observations of 4th and 5th-grade math lessons across 317 teachers in districts serving predominantly historically marginalized students. The dataset features turn-level annotations for dialogic discourse moves paired with rich metadata, including student test scores, demographics, and expert observation ratings. By demonstrating that natural language processing (NLP) models can successfully leverage this data to quantify instructional practices that correlate with student learning outcomes, this open-access resource offers critical infrastructure to study and improve K-12 pedagogy at scale.
Find the paper introducing the dataset here: https://aclanthology.org/2023.bea-1.44/#
Find the GitHub repo with the dataset here: https://github.com/ddemszky/classroom-transcript-analysis
BabbleSim (coming soon)
A Unity-based acoustic classroom simulation environment for generating synthetic audio datasets. BabbleSim places multiple virtual speakers in a 3D classroom, plays real speech and noise recordings through them, and records the resulting spatial audio mix — producing realistic multi-speaker room recordings for speech research.