TutorMoments-Preview

462 real K–12 math tutoring sessions (student and tutor) with human annotations, plus a benchmark of 7,280 AI-tutor attempts scored the same way. A preview release from TutorMoments, a project on how well tutors — human and AI — scaffold, push for rigor, and build rapport. From one K–12 tutoring program (anonymized as tutoring_provider_a).

• Paper: When Help is Unhelpful: Evaluating AI Tutors for Productive Struggle
• Code: https://github.com/allenai/tutormoments
• Contact: Ryan Knight (ryangknight@gmail.com), Lucy Li (lucyli@cs.wisc.edu), Kyle Lo (kyleclo@cs.washington.edu), Kajal Patel (kajal.patel@wisc.edu), Albert Zhang (albert.z.zhang@gmail.com)

Subsets

Config	Rows	Unit
transcripts	462	one tutoring session
annotations	1,584	one annotator pass over a session (263 transcripts; 79 caption / 704 scaffolding / 801 rapport)
ground_truth	207	one session's labeled key moments (10,685 moments: 5,378 scaffolding / 5,307 rapport)
moments	520	one frozen, runnable key moment (260 scaffolding / 260 rigor) — the set the benchmark replays
benchmark_520	7,280	one AI-tutor replay (520 moments × 7 models × 2 prompts)
action_taxonomy	51,246	one decomposed tutor action classified A–M (10,674 human / 40,572 AI tutor; model-generated labels)

from datasets import load_dataset
transcripts = load_dataset("allenai/tutormoments-preview", "transcripts")["train"]

Fields are defined by the bundled JSON Schemas (transcripts.schema.json, annotations.schema.json, ground_truth.schema.json, moments.schema.json, benchmark_520.schema.json, action_taxonomy.schema.json). Common join key: transcript_id (= conversation_id in ground_truth; in moments and benchmark_520 it is the trailing UUID of conv_id, a {tutor_id}_{student_id}_{transcript_id} composite).

moments

The runnable benchmark set: 520 human-annotated key moments (260 scaffolding / 260 rigor), each frozen at its resolved cut point and packaged self-contained for replay — the pre-cut transcript context, the gold dimension, the real human continuation that guides the oracle synthetic student, the frozen student.trait persona, and provenance back to the source transcripts (annotator cut votes, moment ids). This is the set the benchmark harness consumes; benchmark_520 holds its replays (moments.id = {set_name}:{scenario_id}, so it joins to benchmark_520.scenario_id by stripping the balanced_520: prefix). The companion moments.manifest.json carries integrity hashes: content_hash over the canonical records and file_sha256 over moments.jsonl's exact bytes.

benchmark_520

The AI-tutor benchmark behind the paper's leaderboard: from the 520 balanced key moments in moments, each transcript is cut at the moment, a tutor model continues (≤5 turns) with an oracle synthetic student, and the annotator pipeline scores it. 7,280 rows = 520 × 7 tutor models × 2 prompt modes (plain, evaluation-aware). Each row carries the scenario, the AI exchange, and its annotation.

Each row also embeds student_trait — the exact persona its simulated student ran with, recovered from the original run's trait cache with provenance (generator_model, generated_at). All rows for the same scenario share one persona, so every tutor model faced the same students. Re-running the benchmark with these personas reproduces the paper's setup; freshly generated traits would be a different student population.

action_taxonomy

The tutor actions of every scaffolding moment, classified into the 13-category A–M action taxonomy (scheme lm_extended_v1): one row per action_decomposed item, for the human tutors (ground_truth, 10,674 rows over 4,709 moments) and the AI tutors (benchmark_520, 40,572 rows over 7,279 replays). The categories are model-generated: claude-opus-4-8 with provider-native parameters {"thinking": {"type": "disabled"}}, batches of 50, human and AI statements classified in separate pools. Facets removed by the pre-classification filter (statements whose subject is not the tutor, bare stances, and descriptions of what the tutor did not do) are kept as rows with category: null and an excluded_reason; distributions use the classified rows only, as a macro mean over moments. moment_id joins AI rows to moments.id; human rows join to ground_truth by transcript_id (= conversation_id) and key_moment_index.

These labels were regenerated in October 2026 from this dataset at revision 66058b4c7ef5e7631b3c4d39a1dfa8172e36d8bc; they are not the classification run behind the paper. The paper classified a smaller sample — 100 moments per AI cell — so this config does not exactly reproduce the paper's action-distribution figure or its KL-divergence table. The paper's AI cells are recoverable from benchmark_520 by filtering source_round to s42, s43 and topup (100 moments per cell, 50 scaffolding / 50 rigor). The paper's canonical numbers are analysis/working-paper-20260630/v1_action_taxonomy_distribution.csv and kl_divergence_table.tex in the code repository. action_taxonomy.manifest.json records the input revision and file hashes, the classifier configuration and prompt hash, counts, token usage and cost. One benchmark_520 replay has no rows because its scoring failed in the source data (action_decomposed is empty).

Deidentification

Involves minors in an educational setting; treat accordingly.

• IDs (transcript_id, session_id, tutor_id, student_id, annotator_id) are UUIDv5 surrogates.
• The partner program appears only as tutoring_provider_a.
• Free-text PII is replaced with fabricated surrogates or tags (e.g. <Name>); names/emails in the text are not real. Absolute session datetimes are removed.
• Automated and reviewed but imperfect — report residual PII (contact above); do not redistribute affected records.

Limitations

English-dominant with some Spanish; a single program; annotation is subjective; ground_truth labels are LLM-assisted and not infallible. Not for re-identification.

License

This dataset is licensed under CC BY 4.0. It is intended for research and educational use in accordance with Ai2's Responsible Use Guidelines.

Source Attribution

The replay data includes synthetic tutor/student conversations between LMs generated from third-party models, which are subject to the respective model providers' terms and conditions. We used the following models to generate the replay data: Claude Opus 4.8 and Sonnet 4.6 (provided by Anthropic); DeepSeek v4 Pro (provided by Hangzhou DeepSeek Artificial Intelligence Co.), Gemini 2.5 Pro and 3.5 Flash (provided by Google), and GPT 5.5 and 5.4 mini (provided by OpenAI). The action_taxonomy categories were generated with claude-opus-4-8 (provided by Anthropic).

Acknowledgements

This project has been made possible in part through support from the Gates Foundation and Learning Commons.

Citation

A paper is in preparation; this citation will be updated when it is published. Until then, cite the dataset directly:

@misc{zhang2026whenhelp,
  author       = {Zhang, Albert and Ross, Alexis and Patel, Kajal and Bernado, Julian and
                  Bowie, Rebecca and Trindade Ribeiro, Ana and Halper, Daniel and
                  Valayaputtur, Haripriya and Andreas, Jacob and Loeb, Susanna and
                  Lucy, Li and Lo, Kyle and Knight, Ryan},
  title        = {When Help is Unhelpful: Evaluating AI Tutors for Productive Struggle},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/allenai/tutormoments-preview}}
}