juiceb0xc0de/jev-canary-09-26

Jev History Corpus: a post-cutoff knowledge canary

A synthetic question-answer corpus about the Jev family of decision models, built as a diagnostic for fine-tuning pipelines. The subject matter postdates September 2026, so a model with an earlier pretraining cutoff should know almost none of it. Train on the corpus, score the held-out questions before and after, and the difference shows whether a training run is writing new knowledge into the weights.

Version 1.0 contains 10,389 training pairs and 265 held-out evaluation questions, drawn from 58 source units and 2,616 extracted fact records.

Intended use

The corpus is a test of the training setup, not a knowledge resource in its own right. It answers one question: did the training land? Typical uses:

• validating a new or modified training method
• comparing LoRA with full fine-tuning, or sweeping learning rates, on a target that cannot be confused with prior knowledge
• checking a data loader, chat-template handling, or packing logic after a rewrite
• smoke-testing a knowledge-injection pipeline before committing a larger dataset to it

Why Jev

The Jev family has a well-defined start date. TypeSafe's launch post is dated September 15, 2026, and the papers the corpus draws on carry arXiv identifiers from September and October 2026 (the 2609 and 2610 series). The latest dated source in the corpus is a usage guide from November 14, 2026. Coverage includes TypeSafe's hosted Jev models, Open-Jev, the JEV-9B, JEV-27B, JEV-27B-VL and JEV-35B checkpoints, GEV-26B-Decide, community derivatives such as NeoHorse-Jev-4B and Jeva-4B, and the surrounding benchmarks, model cards, documentation, GitHub repositories and Hub discussions.

A model whose pretraining ended before September 2026 should score near zero on the held-out set before training.

A date is not proof of absence. The Jev sources cite earlier work (2024 and 2025 papers), build on earlier base models such as Qwen3.5, use common benchmarks such as MMLU-Pro, and at least one source repository predates the September 2026 burst, so a model can earn partial credit by reasoning from general background. Measure the baseline on the eval set before training. The canary is informative when the baseline is low and the post-training score is clearly higher.

Dataset structure

File	Rows	Contents
jev_qa_train.jsonl	10,389	training pairs, single-turn chat format
jev_qa_train.meta.jsonl	10,389	per-row provenance for the training pairs, row-aligned with the file above
jev_heldout-v0-1.jsonl	265	evaluation questions with gold answers and grading keywords
build_report.json		build summary: row counts, per-unit and per-type counts, exclusions

Training rows

Each row is a single-turn chat:

{"messages": [{"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}]}

All 10,389 questions and all 10,389 answers are unique; 15 duplicate questions were removed during the build. Questions have a median length of 12 words (range 3 to 31) and answers a median length of 51 words (range 9 to 139). No evaluation question appears in the training set.

Question types:

Type	Rows
numbers	1,852
comparison	1,575
definition	1,387
limitation	1,253
mechanism	1,248
recipe	974
why	958
timeline	614
unknown	528

Training metadata

Row i of jev_qa_train.meta.jsonl describes row i of jev_qa_train.jsonl:

Field	Contents
unit	id of the source unit the pair was written from
qtype	question type, as in the table above
facts	indices of the fact records, within that unit's training facts, that the pair was written from
facts_file	path of the unit's training-fact file during the build; the fact base itself is not part of this release

Each pair draws on between 1 and 12 fact records: 4,622 pairs draw on one, 3,720 on two, 1,487 on three, and 560 on four or more. Every training fact record in every unit is covered by at least one pair.

Evaluation rows

Field	Contents
question	the held-out question
gold	the reference answer
keywords	one to four terms a correct answer is expected to contain
meta	unit, fact and facts_file for the held-out fact the question was written from

There is one question per held-out fact. Evaluation questions have a median length of 20 words and gold answers a median length of 28 words; 56 of the 58 source units contribute held-out questions. Grade against gold, by hand or with an LLM judge. The keywords are a screening aid for a quick first pass and are not sufficient on their own as an automated correctness check.

Loading

from datasets import load_dataset

train = load_dataset("juiceb0xc0de/jev-history-corpus", "train")["train"]
meta = load_dataset("juiceb0xc0de/jev-history-corpus", "meta")["train"]
heldout = load_dataset("juiceb0xc0de/jev-history-corpus", "eval")["test"]

Sources

58 source units, identified by the unit field:

Unit prefix	Units	Material
paper_	10	arXiv papers 2609.25845, 2609.26758, 2609.34736, 2610.02486, 2610.05107, 2610.06425, 2610.06744, 2610.17309, 2610.18773, 2610.30970
eq_	10	equation extracts from the same ten papers
docs_	12	documentation pages, in blocks of sixteen
cards_	7	model cards, in blocks of ten, plus a catch-up block
repo_	11	GitHub repositories, named repo_<owner>_<name>
discussions_	3	Hub discussions
issues_	2	GitHub issues
faq, blog_typesafe_launch, census_hub	3	a FAQ, TypeSafe's launch blog post, and a census of Jev-related Hub repositories

Construction

1. The sources were consolidated into a local fact base of 2,616 fact records across 58 units.
2. Within each unit, the facts were split at the fact-record level: 274 were withheld and 2,342 kept for training. Nine withheld facts that shared a long quote and a source with training material were excluded (listed in build_report.json), leaving 265 held-out facts.
3. The evaluation set was written first, before any training pair: one question per held-out fact, 265 in all, each checked against its fact. It has not changed since.
4. Every later round wrote training pairs from the 2,342 training facts only. Claude Opus 5.5 (claude-opus-5-5) wrote each unit's pairs from that unit's training facts; the writer never saw full source digests or held-out facts. Each batch received a separate checking pass confirming that every pair matches the facts it was written from.
5. Generation produced a checked pool of 11,575 pairs; after the checking pass and removal of 15 duplicate questions, 10,389 remain.

Evaluation protocol

1. Score the untrained model on the eval set and record the baseline.
2. Train on the training rows with the method under test.
3. Score the eval set again. A clear rise over the baseline indicates that training is writing new knowledge; a flat result indicates a problem somewhere in the pipeline.
4. Run a general-capability benchmark before and after, so that gained knowledge can be told apart from damage to the base model.

Limitations

• The train/eval split is by fact record, not by concept. Obvious shared text and long quotes were removed from the evaluation set, but shared concepts and semantic overlap across sources can remain. Treat this as a diagnostic canary rather than a certified unseen-knowledge benchmark.
• The corpus records what the sources say, not whether the sources are correct. Checking confirms that each pair matches the facts it was written from; it does not independently verify the sources' claims.
• Source coverage is partial. The corpus does not cover everything published about Jev.
• Every row is synthetic, written by a model from extracted facts. No row was written or reviewed by a human annotator.
• The post-cutoff property depends on the model under test. A model trained on data from after September 2026 may already have seen Jev material, which is why the baseline measurement is part of the protocol.

Citation

@misc{holmberg2026jevcorpus,
  author = {Holmberg, Rick},
  title = {Jev History Corpus: a post-cutoff knowledge canary},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/juiceb0xc0de/jev-history-corpus}
}