Data sources
Every dataset the book draws on, with its years, weights and license, and the chapters that use it.
Amazon Reviews 2023
Amazon Reviews 2023, Grocery and Gourmet Food, coffee products (McAuley Lab)
- Id
- amazon-reviews-2023
- License
- No license stated; the booklet shows aggregates and short excerpts only
Cite as: Hou, Y., Li, J., He, Z., Yan, A., Chen, X., & McAuley, J. (2024). Bridging language and items for retrieval and recommendation. arXiv:2403.03952.
Reviews come from people who chose to write them. The coffee subset is selected by a written rule over category paths and product titles.
Used in From evidence to typed questions, From probabilities to actions, Comparing decision systems.
CFPB Consumer Complaint Database
Consumer Financial Protection Bureau, Consumer Complaint Database, narratives filed under the August 2023 product taxonomy (copy retrieved July 2026)
- Id
- cfpb-complaints
- License
- Public domain; the booklet shows aggregates only
Cite as: Consumer Financial Protection Bureau. Consumer Complaint Database.
Consumers choose the product and issue themselves. Narratives were published only with consent and after scrubbing, and the database is not a statistical sample of consumer problems. The bureau stopped publishing narratives on August 14, 2026; this booklet uses a copy retrieved in July 2026 that keeps the first 3,000 characters of each narrative.
Used in Resolving and routing customer complaints, Comparing decision systems.
Constructed agent tasks
Short agent tasks with completion criteria, graders and fault-injected traces, written for this booklet
- Id
- constructed-agent-tasks
- License
- Authored for this booklet
Cite as: Singh, V. (2026). Constructed agent tasks for Decision AI.
The mix of injected faults describes the test harness, not a production population.
Used in Deciding when an agent is finished, Comparing decision systems.
Constructed patches
Code changes written for this booklet, each with a test that detects its defect or confirms it is acceptable
- Id
- constructed-patches
- License
- Authored for this booklet
Cite as: Singh, V. (2026). Constructed patches for Decision AI.
Constructed examples are reported separately from sampled ones.
Used in Allocating review attention to code changes, Comparing decision systems.
FEVER
Fact Extraction and VERification claims with Wikipedia evidence
- Id
- fever
- License
- CC BY-SA 3.0
Cite as: Thorne, J., Vlachos, A., Christodoulopoulos, C., & Mittal, A. (2018). FEVER: a large-scale dataset for fact extraction and verification. NAACL.
Gold evidence is incomplete.
No chapter uses this dataset yet. Planned
HuffPost News Category Dataset
News Category Dataset v3: HuffPost headlines and short descriptions, 2012 to 2022
- Id
- huffpost-news
- License
- CC BY 4.0
Cite as: Misra, R. (2022). News Category Dataset. arXiv:2209.11429.
Most rows predate May 2018.
No chapter uses this dataset yet. Planned
Open Food Facts
Open Food Facts product database
- Id
- open-food-facts
- License
- ODbL 1.0 (database), DbCL 1.0 (contents), CC BY-SA 3.0 (images)
Cite as: Open Food Facts contributors. Open Food Facts product database.
NOVA groups are computed by a rule over parsed ingredients.
No chapter uses this dataset yet. Planned
Routing task family
A task family with checkable answers, chosen by a stated pilot rule
- Id
- routing-tasks
- License
- Depends on the family chosen
Cite as: Chosen in Release 2.
Chosen in Release 2 by the pilot rule in the build guide.
No chapter uses this dataset yet. Planned
SWE-bench Verified
SWE-bench Verified, 500 human-validated issues from 12 Python repositories
- Id
- swe-bench-verified
- License
- No license on the dataset card; repositories carry their own licenses
Cite as: Jimenez, C. E., et al. (2024). SWE-bench: Can language models resolve real-world GitHub issues? ICLR.
A 2026 audit reported flawed tests on a share of hard tasks, so a failing patch is not necessarily defective.
No chapter uses this dataset yet. Planned
SWE-rebench
SWE-rebench software engineering tasks (Nebius)
- Id
- swe-rebench
- License
- CC BY 4.0, plus the license of each source repository
Cite as: Nebius (2025). SWE-rebench.
Tasks are extracted continuously from public repositories.
Used in Allocating review attention to code changes, Comparing decision systems.
SWE-rebench OpenHands trajectories
Agent trajectories and patches on SWE-rebench tasks from one model on one scaffold (Nebius)
- Id
- nebius-openhands-trajectories
- License
- CC BY 4.0
Cite as: Nebius (2025). SWE-rebench OpenHands trajectories.
Patches come from one agent configuration, not from every agent.
Synthetic resumes
Resumes generated from seeded templates in matched sets that differ only in name
- Id
- synthetic-resumes
- License
- Authored for this booklet
Cite as: Singh, V. (2026). Synthetic resumes for Decision AI.
No real people.
No chapter uses this dataset yet. Planned