Datasets:

chaoliangUNSW
/

jev-style-2b-v4-training

Exception:    SplitsNotFoundError
Message:      The split names could not be parsed from the dataset config.
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 286, in get_dataset_config_info
                  for split_generator in builder._split_generators(
                                         ~~~~~~~~~~~~~~~~~~~~~~~~~^
                      StreamingDownloadManager(base_path=builder.base_path, download_config=download_config)
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  )
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/webdataset/webdataset.py", line 84, in _split_generators
                  raise ValueError(
                  ...<2 lines>...
                  )
              ValueError: The TAR archives of the dataset should be in WebDataset format, but the files in the archive don't share the same prefix or the same types.
              
              The above exception was the direct cause of the following exception:
              
              Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/config/split_names.py", line 68, in compute_split_names_from_streaming_response
                  for split in get_dataset_split_names(
                               ~~~~~~~~~~~~~~~~~~~~~~~^
                      path=dataset,
                      ^^^^^^^^^^^^^
                      config_name=config,
                      ^^^^^^^^^^^^^^^^^^^
                      token=hf_token,
                      ^^^^^^^^^^^^^^^
                  )
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 340, in get_dataset_split_names
                  info = get_dataset_config_info(
                      path,
                  ...<6 lines>...
                      **config_kwargs,
                  )
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 291, in get_dataset_config_info
                  raise SplitsNotFoundError("The split names could not be parsed from the dataset config.") from err
              datasets.inspect.SplitsNotFoundError: The split names could not be parsed from the dataset config.

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

Jev-Style 2B Decision v4 training pool

The complete training pool of Jev-Style-2B-Decision-v4 (text and image decisions with a calibrated probability for every option), published so that the model card's Training data and licences section can be checked row by row and the run can be rebuilt. It is a mixed-source dataset: every row carries the licence of its own source (see Licences), and the pool is not released under one permissive licence.

Related: Jev-Style-2B-Decision-v4 (the model trained on it; also GGUF and MLX builds) · decision-index-results (the model's Decision Index runs) · v4 collection · jevstyle.com.

Files in this repository

File	Size	Contents
`pool-main.tar`	7.77 GB	The assembled pool: `data/pool-v4/main/{train,dev,cal}.jsonl`, `quotas.json`, `manifest.json`; the training images under `data/images/`; the run configs under `configs/v4/`. Uncompressed tar.
`code.tar.gz`	2.8 MB	A snapshot of the training workspace at upload time (2026-10-25): `src/` (the `macjev` training, rendering, readout and evaluation code; `cloud_v3`), `scripts/` (pool assembly, teacher jobs, gates, uploads), `configs/v4/`, `docs/` (data contract, runbooks, plans) and the working notes (`STATE.md`, `LOG.md`, `DECISIONS.md`, `CLAUDE.md`).
`code/code-2026-10-25.tar.gz`	4.8 MB	An earlier snapshot of the same workspace, kept for reference.
`SHA256SUMS`		sha256 of `pool-main.tar` and `code.tar.gz`.

pool-main.tar was packed on macOS and also contains AppleDouble entries (._main, ._train.jsonl, ...). They are harmless metadata files; skip them with tar xf pool-main.tar --exclude '._*'.

Inside `pool-main.tar`

data/pool-v4/main/train.jsonl   916,923 rows   (4.1 GB)
data/pool-v4/main/dev.jsonl      59,995 rows   (278 MB)   held out: checkpoint selection
data/pool-v4/main/cal.jsonl       3,325 rows   (11 MB)    held out: the per-question-type temperatures
data/pool-v4/main/quotas.json                             per-component token quotas used by the run
data/pool-v4/main/manifest.json                           counts per source and split, gate fingerprints, drop reasons
data/images/{aokvqa_train,vqav2_train,textvqa_train,cord_train,scienceqa_train,smoke}/<sha256>.png
configs/v4/                                               the training configs

The images are the training copies of the A-OKVQA and VQAv2 items (COCO photos), TextVQA (Open Images photos), CORD receipts and ScienceQA figures, at the resolution they were trained at (about 3.3 GB in total); smoke/ is a few rows used for the pre-flight checks.

Row format

One JSON object per line, the v3 format plus the v4 extensions (docs/DATA_CONTRACT_v4.md in code.tar.gz is the authoritative description):

{id, group_id, split, source, category, lang, license,
 state, question: {t: "choice" | "noul" | "score", ins, crit},
 label: {valid, soft}, pieces,
 meta: {component, family, label_method, data_class, overlap, images?}}

state is the text or JSON the model reads; image rows mark the position with <|image_k|> and list the files under meta.images (path, grid_thw, merge, n_tokens, from macjev.vision.prepare_image).
question.t is the question type (choice with named options in crit, noul yes/no, score over ordered levels), ins the instruction and crit the options or levels.
label.valid is the gold answer; label.soft is a probability distribution over the options where one exists.
pieces holds the pre-rendered token ids (Qwen3.5 tokenizer), so the pool trains without re-rendering.
license is the licence recorded for the row's source; meta.data_class and meta.label_method say where the row and its label came from (official train split, code generator, teacher model, Jev soft label, ...).

Soft labels distilled from Jev 1.14

334,739 of the training rows carry Jev 1.14's probabilities as their target (label.soft), obtained through TypeSafe's public API (/v1/systemone). Where Jev's top choice was not a correct answer, the gold label was kept instead (58,586 rows). No Jev weights or code are included. JevStyle is an independent project with no TypeSafe affiliation or endorsement; the model card discloses this use in full.

Licences

license: other because the pool mixes sources with different terms. The model card's Training data and licences section and the model repository's validation/data_sources.json list all 88 sources with the licence recorded for each; the per-row license field applies the same information to each row. In short:

Official benchmark train splits (Apache-2.0, MIT, CC BY 2.0/3.0/4.0, CC BY-SA 4.0): BANKING77, CLINC150, SGD, ContractNLI, ARC, WinoGrande, HellaSwag (ActivityNet-caption items only), GSM8K, POP909, Amazon ESCI, iSarcasmEval, Habermas Machine, CommonsenseQA, RAGTruth, HoVer, When2Call, New Yorker caption contest, HelpSteer2/3, WANLI, TabFact, DROP, BoolQ, SQuAD v2, SNLI, MASSIVE, CORD and the other sources named on the model card.
Share-alike sources (CC BY-SA): SQuAD v2, BoolQ, DROP, SNLI, ARC, SGD, HoVer, ShARC, TempReason, ScienceQA.
Sources with terms of their own: jailbreak prompts from the jailbreak_llms collection (research purposes only), the research-only rows of neuralchemy's prompt-injection set were removed; HellaSwag's upstream repository is blocked after a DMCA notice. Check each source's terms before commercial use.
Images: A-OKVQA and VQAv2 use COCO photos (Flickr, under their own licences and the COCO terms of use); TextVQA uses Open Images photos (CC BY 2.0); CORD (CC BY 4.0); ScienceQA (CC BY-SA 4.0, mirror).
Outputs of other models: the Jev 1.14 soft labels above; documents, questions, reviews and stories written by Anthropic Claude and OpenAI GPT models (labels from code, from the source data, or from two models that agreed); third-party sets that contain model-written text (WANLI, HelpSteer2/3, When2Call, RAGTruth, part of the jackhhao benign prompts). The providers' terms of use may restrict how models trained on such outputs may be used.
Decision Index hygiene: every row was checked against the public Decision Index 0.3 suite (row hashes, 8-gram and MinHash gates; 19,387 rows dropped); no suite rows and no suite build scripts are in the pool. The pool still contains the official train splits of several benchmarks the index tests on, so results on those are not zero-shot.

Code in code.tar.gz is Apache-2.0 (the same licence as the model); the working notes it contains are informal project notes, kept as they were at upload time.

Download

pip install -U huggingface_hub
hf download chaoliangUNSW/jev-style-2b-v4-training --repo-type dataset --local-dir jev-v4-training
cd jev-v4-training && sha256sum -c SHA256SUMS
tar xf pool-main.tar --exclude '._*'     # -> data/pool-v4/main, data/images, configs/v4
tar xzf code.tar.gz                      # -> src, scripts, configs, docs

docs/H100_RUNBOOK.md in code.tar.gz is the step list the training run followed.

Disclaimers

Jev-Style is not affiliated with, endorsed by or connected to TypeSafe AI or Jev. Part of the training rows use soft labels distilled from TypeSafe Jev 1.14 via its public API; no Jev weights or code are included. Not affiliated with the Laya authors or the Qwen team. Maintained by Chaoliang Yan.