JevEmbed-Data: sources and checks
This release has 1,667,639 questions: 453,880 Choice, 627,933 Score, and 585,826 Noul. It contains 1,601,157 training questions and 66,482 validation questions. The Hub calls the validation split test; it is not a separately collected test set.
There are 490,242 train groups and 16,240 validation groups. A group may hold several questions about one case or response, so question counts are larger than counts of independent cases.
Sources
The table counts final questions, not unique prompts. Each Parquet row also records its source, revision, row reference, and license.
| Source | Questions | How it was used | License |
|---|---|---|---|
| procedural-jev | 399,000 | Original Jev decisions; rules checked. Added true/false wording to Noul without changing labels. | Apache-2.0 |
| PKU-SafeRLHF | 313,660 | Human safety labels → Noul/Score; safe-response preferences → Choice. | CC-BY-NC-4.0 |
| Open-Jev | 276,249 | Native Choice, Score, Noul, and soft labels. | CC0-1.0 |
| Prometheus Feedback Collection | 99,668 | GPT-4 ratings with reference answers → Score. | CC-BY-4.0 |
| HelpSteer2 | 89,305 | Human 0–4 ratings → Score. | CC-BY-4.0 |
| UltraFeedback | 86,146 | Selected GPT-4 ratings → Score; one rating per completion. | MIT |
| HH-RLHF | 69,759 | Helpful pairs → Choice; harmless pairs → Noul. | MIT |
| HelpSteer | 69,655 | Human 0–4 ratings → Score. | CC-BY-4.0 |
| Nectar | 65,497 | One top-versus-bottom Choice per ranked prompt. | Apache-2.0 plus card restriction |
| Jev Distill Corpus v3 | 38,540 | Selected high-confidence teacher Choice/Noul labels. | Apache-2.0 |
| Open-Jev v1.1 | 35,840 | Original community decision questions. | CC0-1.0 for selected rows |
| LMSYS Arena | 29,508 | Clear single-turn human preferences → Choice. | Apache-2.0 |
| HelpSteer3 | 27,129 | Strong preferences → Choice; unanimous human feedback → Score. | CC-BY-4.0 |
| BeaverTails | 22,106 | Human safety labels → Noul. | CC-BY-NC-4.0 |
| JevEmbed-generated cases v2 (synthetic) | 18,000 | 8,000 cases on refunds, invoices, Python, SQL, and incidents; answers recalculated or executed. | Apache-2.0, project-original |
| System One 270M | 13,217 | Consistent high-confidence teacher Choice/Score/Noul labels. | Apache-2.0 |
| SystemOne Lite Phase 2 | 6,140 | Budget decisions with independently checked answers. | Apache-2.0 |
| VEJI-Synth v2 | 5,220 | Date and invoice cases whose answers could be recalculated. | CC0-1.0 |
| JevEmbed-generated cases v1 (synthetic) | 3,000 | 600 cases each for Python execution, SQL queries, dates, unit conversion, and probability; answers checked by execution or calculation. | Apache-2.0, project-original |
The corpus has mixed licenses. The license shown for each row belongs to that source; there is no single license for the whole release. Nectar's card adds a restriction beyond Apache-2.0. Check source terms before reuse or redistribution.
Preparation and checks
Native Jev questions retain their Choice, Score, and Noul targets, including soft probabilities. Human ratings become Score labels, preference comparisons become Choice labels, and safety judgments become Noul labels. Teacher labels remain traceable through each row's source fields. The two JevEmbed-generated sets are synthetic: their Python and SQL answers were executed, and their numeric and rule-based answers were recalculated.
Published validation splits were preserved for Open-Jev, procedural-jev, HelpSteer, HelpSteer2, and the selected VEJI cases. For other sources, a fixed hash of the case ID assigns about 2% of groups to validation. Questions from the same case stay together. Text was normalized and exact state/question duplicates were resolved across sources.
Soft probability labels remain probabilities. The rendered embedding inputs fit within 2,048 tokens under the KaLM-Embedding-V2.5 tokenizer. Other tokenizers may produce different lengths.
| Check | Result |
|---|---|
| JevEmbed request and training-label validation | 1,667,639 rows passed; 0 errors. |
| Exact input and group overlap between train and validation | 0. |
| Procedural Noul descriptions versus original state, question, and label | 168,000 passed. |
| JevEmbed-generated cases v2 | 18,000 questions from 8,000 cases passed independent rule, Python, and SQL checks. |
The five largest sources provide 70.6% of questions, so source sizes are uneven. Human, teacher, and synthetic labels are identified by their source rather than treated as interchangeable.
The upload's MANIFEST.json lists each Parquet file, row count, size, and SHA-256 checksum. The original source revision and license are also stored on every row.