JEVal

General Decision Models:
Benchmarking and Insights Beyond Jev

Paper  |  JEVal Dataset  |  GitHub  |  PDF  |  Citation

InnerJev-4B: Full / LoRA  |  InnerJev-27B: Full / LoRA

---

JEVal benchmarks general decision models: models that return structured decisions and probabilities over candidate options. It unifies 36 datasets, 58 subtasks, and 11,257 evaluation examples across 10 application domains, with 9,457 English and 1,800 Chinese examples.

This release stores all examples in one Parquet file and one test split. The 58 original subtask identifiers remain in the source column. Input texts, instructions, option order, empty values, question IDs, labels, and original sample identifiers are preserved without filtering or resampling.

Examples	Datasets / subtasks	Domains	Question formats
11,257	36 / 58	10	9,582 choice · 1,356 noul · 319 score

JEVal overview (Figure 2 in the paper PDF): construction of the unified decision interface, coverage of ten domains, and input-length distribution. This distribution uses cl100k_base tokenization of the state, instruction, and candidates.

InnerJev models

We release InnerJev at 4B and 27B scales, with full-rank checkpoints and LoRA adapters. The First-Token Decision Readout obtains probabilities from the logits of valid option-label tokens in a single forward pass, without generating an answer sequence. Reasoning-to-Readout Self-Distillation trains this readout to match the same model's answer distribution after reasoning.

Model	Base model	Full-rank checkpoint	LoRA adapter	Training data
InnerJev-4B (ours)	Qwen3.5-4B	InnerJev-4B-Full	InnerJev-4B-LoRA	4B training data
InnerJev-27B (ours)	Qwen3.8-27B	InnerJev-27B-Full	InnerJev-27B-LoRA	27B training data

See the InnerJev code repository and individual model cards for inference and adapter-loading instructions.

Load the dataset

import json
from datasets import load_dataset

dataset = load_dataset("carlosxiang/JEVal", split="test")
row = dataset[0]
native = {
    "source": row["source"],
    "state": row["state"],
    "questions": json.loads(row["questions_json"]),
    "label": json.loads(row["label_json"]),
}
# Only state and questions enter a model request; label is used for scoring.

The repository contains one default configuration and one test split. Use the Dataset Viewer to inspect rows and filter by source or question_type. For a local copy, use load_dataset("parquet", data_files={"test": "data/test.parquet"}, split="test").

Main results

The table below reproduces Table 1 of arXiv v1 for all 25 model configurations. Classification metrics cover the 10,938 choice and noul questions; the 319 score questions are evaluated separately in the paper. Accuracy counts invalid responses as incorrect. ECE and raw Brier score use valid responses. Higher accuracy and lower ECE/Brier are better.

Model	Mode / reasoning effort	ACC (%) ↑	ECE ↓	Brier (raw) ↓
Jev	Decision	78.38	0.0532	0.287
Kev-0.8b	Decision	49.63	0.0432	0.617
Kev-4b	Decision	68.90	0.0843	0.435
Kev-9b	Decision	70.87	0.0579	0.404
Kev-27b	Decision	77.17	0.0356	0.304
Intern-Decision-0.8B	Decision	47.71	0.0635	0.643
Intern-Decision-2B	Decision	55.72	0.0214	0.553
Intern-Decision-4B	Decision	70.74	0.0152	0.399
Laya	Decision	26.30	0.2485	0.833
Openjev-4b-v5	Decision	70.20	0.1472	0.481
Nimble-9b	Decision	73.87	0.1045	0.369
Nanojev	Decision	25.34	0.1300	0.806
Lightjev	Decision	37.20	0.1145	0.742
Jevforge-0.8b	Decision	31.04	0.1164	0.794
InnerJev-4B (ours)	Decision	71.92	0.0843	0.383
InnerJev-27B (ours)	Decision	79.24	0.0149	0.279
Qwen3.5-4B	Non-thinking	61.37	0.2526	0.781
Qwen3.8-27B	Non-thinking	73.81	0.1697	0.446
DeepSeek-V4-Flash	Non-thinking	72.83	0.1923	0.494
GPT-6-Luna	Non-thinking	80.93	0.1159	0.297
GPT-6-Sol	Non-thinking	79.92	0.1309	0.322
Qwen3.5-4B	Thinking / default	70.73	0.1716	0.489
Qwen3.8-27B	Thinking / xhigh	79.49	0.1005	0.319
DeepSeek-V4-Flash	Thinking / high	79.69	0.1449	0.382
GLM-5.3-Flash	Thinking / high	80.37	0.0784	0.316

Bold metric values mark the best result in each column across the configurations shown. These are the published Table 1 results, not a new evaluation run.

Results by application domain

Figure 3 from the paper PDF. Domain-level accuracy, ECE, and raw Brier score for the evaluated configurations. N/T denote non-thinking/thinking modes. Domains: general knowledge and reasoning (GKR), medicine (MED), law (LAW), finance (FIN), commonsense (CS), agent and tool use (ATU), factuality and hallucination (FH), long-context and document understanding (LCDU), social science (SS), and personalization and preferences (PP).

Latency versus input length

Figure 4 from the paper PDF. Mean latency of valid responses on a shared 200-question sample, grouped into six input-length bins up to 16K Jev-reported tokens. The vertical axis is in seconds on a logarithmic scale. Local models were measured on RTX 4090 GPUs; their processing times exclude queueing and network transport. The hosted Jev curve measures HTTP latency and includes network/provider queueing. Laya retains its 8K input limit and official truncation. These timing boundaries differ and should be considered when interpreting the comparison.

Building InnerJev

Figure 9 from the paper PDF. Left: restrict first-answer-token logits to valid option labels, then return a selected option (choice), the probability of true (noul), or an expected rating (score). Right: distill the same model's post-reasoning answer distributions into its direct readout, with a frozen-model KL anchor.

Repository files

• data/test.parquet: the complete merged evaluation dataset.
• scripts/merge_eval.py: the source merge and lossless validation script.
• README.md: paper overview, model links, published results, schema, loading instructions, and source inventory.
• assets/: four figures extracted from the published arXiv v1 source.
• citation.bib: the recommended arXiv BibTeX citation.

Columns

Column	Meaning
id	Unique row ID in this frozen merge: source name and release row position.
source	Original source/subtask identifier, unchanged (58 distinct values).
state	Complete original input state text.
question_type	choice, noul, or score.
questions_json	Lossless JSON encoding of the complete original questions dictionary.
label_json	Lossless JSON encoding of the original label, retaining its value and type.
source_item_id	Original logical identifier: source dataset / source split / zero-based source row index.

questions_json and label_json use JSON text because the native records contain heterogeneous dictionary/list/null structures and string/numeric labels. Applying json.loads restores the native values.

Source traceability

All rows retain original item identifiers containing their source splits and source positions. For example, source_item_id = ai2-arc-challenge/test/218 identifies a row in the original formal dataset: the test split, at zero-based position 218. This source position is distinct from its position in the merged JEVal table. Original detailed provenance remains in the input tree and is used to validate the merge.

Dataset-level source files, builder versions, sampling seeds, selected indices, exclusion guidance and grouping notes are embedded in the Parquet schema metadata under jeval.sources. When rebuilding from the original inputs, the merge script also writes a local _build/manifest.json. The public layout uses dataset provenance columns rather than date-based directories. Historical dates in source provenance remain part of the audit trail and sampling seed; they do not define separate JEVal subsets.

The JEVal test split denotes this evaluation collection. The upstream formal split is retained inside source_item_id; it may be test, dev, or train. Sampling follows the original frozen inputs and may fragment groups from the upstream dataset, as recorded in the original grouping metadata.

Reproduce and verify the merge

python scripts/merge_eval.py --source-root /path/to/jev/eval --output /path/to/new/JEVal

Requires Python and PyArrow. The builder refuses to overwrite existing output files unless --replace is explicitly supplied and refuses any destination inside the source eval tree. It verifies every native record against the original provenance hash, writes one Parquet file, reads it back, and checks exact native values, option order, label types and source identifiers. It also compares all input files' checksums, sizes and modification times before and after the merge. When rebuilding from the original inputs, validation is recorded locally in _build/validation.json, and _build/input_snapshot.json records the source snapshot. Those build records are not included in this repository. The Parquet file contains evaluation inputs and labels, not model predictions. Published aggregate results and figures are included in this dataset card; per-request prediction logs are not included.

Source inventory

The 58 identifiers below correspond to the benchmark subtasks; some share an underlying source dataset, giving 36 datasets under the paper's grouping.

Source	Examples	Question types
agentharm	200	score: 103, noul: 97
ai2-arc-challenge	200	choice: 200
ai2-arc-easy	200	choice: 200
aida-testc	200	choice: 200
bbq	200	choice: 200
boolq	200	noul: 200
ceval	200	choice: 200
clinc150	200	choice: 200
cmb	200	choice: 200
cmexam	200	choice: 200
cmmlu	200	choice: 200
commonsense-qa	200	choice: 200
communitybench-com-pred	200	choice: 200
communitybench-pref-id	200	choice: 200
compred-finance	200	choice: 200
compred-gender_sexuality	200	choice: 200
compred-history	200	choice: 200
compred-politics	200	choice: 200
compred-science	200	choice: 200
cuad	200	noul: 200
decider	200	choice: 164, noul: 36
disc-law-multi	200	choice: 200
disc-law-single	200	choice: 200
fever	200	noul: 200
financebench-yes-no	37	noul: 37
finqa-yes-no	22	noul: 22
global-opinions	200	choice: 200
go-emotions	200	choice: 200
gpqa	198	choice: 198
halueval-dialogue	200	choice: 200
halueval-general	200	noul: 200
halueval-qa	200	choice: 200
halueval-summarization	200	choice: 200
hellaswag	200	choice: 200
helmet	200	choice: 200
jevbench	200	noul: 66, choice: 118, score: 16
jevforge	200	choice: 102, noul: 98
lex-glue-case-hold	200	choice: 200
lex-glue-ecthr-a	200	choice: 200
lex-glue-ecthr-b	200	choice: 200
lex-glue-ledgar	200	choice: 200
lex-glue-scotus	200	choice: 200
lex-glue-unfair-tos	200	choice: 200
longbench-lsht	200	choice: 200
longbench-retrieval-en	200	choice: 200
longbench-retrieval-en-e	200	choice: 200
longbench-retrieval-zh	200	choice: 200
longbench-trec	200	choice: 200
longbench-trec-e	200	choice: 200
med-qa	200	choice: 200
med-qa-en	200	choice: 200
mmlu-pro	200	choice: 200
nlu-eval-68	200	choice: 200
prism-alignment	200	choice: 200
prism-alignment-score	200	score: 200
scifact	200	noul: 200
truthfulqa	200	choice: 200
winogrande	200	choice: 200

Attribution and licensing

The source dataset names and original source identifiers are retained. Upstream dataset URLs, citations and license terms have not yet been consolidated into a single release license. No blanket license is assigned to third-party data here. Additional source citations and licensing metadata remain to be documented.

Citation

If you use JEVal, InnerJev, or the findings in this work, please cite our paper:

General Decision Models: Benchmarking and Insights Beyond Jev
Feiyu Duan, Jiayu Lin, Jia Wang, Jun Xiang, Jialiang Wu, Xinnong Zhang, Hanqi Yan, Siyuan Wang, and Zhongyu Wei.
arXiv:2610.03935, 2026.

@misc{duan2026generaldecisionmodelsbenchmarking,
      title={General Decision Models: Benchmarking and Insights Beyond Jev}, 
      author={Feiyu Duan and Jiayu Lin and Jia Wang and Jun Xiang and Jialiang Wu and Xinnong Zhang and Hanqi Yan and Siyuan Wang and Zhongyu Wei},
      year={2026},
      eprint={2610.03935},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2610.03935}, 
}

Download the BibTeX citation. Please also respect the attribution and licensing requirements of the underlying source datasets.