JonusNattapong/jev-my-bro-dataset

jev-my-bro Governance Dataset

Provenance-aware English/Thai dataset for training and evaluating the typed jev-my-bro decision model. Each case asks four structured governance questions about an operation:

• action: execute, ask_user, or reject
• needs_review: whether explicit human review/approval is required
• prohibited: whether the operation should be prohibited
• risk: five-level operational risk

This snapshot contains 8,508 cases / 34,032 typed decisions.

This is research/development data, not a production authorization policy. Imported labels marked rule_reviewed are rule-derived and consistency reviewed; they are not human annotations.

Files and splits

Split	Cases	Decisions	English	Thai
train	5,731	22,924	3,829	1,902
validation	879	3,516	586	293
calibration	944	3,776	641	303
test	954	3,816	616	338
Total	8,508	34,032	5,672	2,836

The four splits are kept independent. Cases derived from the same scenario_family stay in one split to prevent family/paraphrase leakage.

Source composition

Source	Cases	Language	Upstream license
jev-my-bro static bootstrap	1,008	EN + TH	Apache-2.0
AmazonScience/massive (th-TH)	2,500	TH	CC BY 4.0
mteb/banking77	2,500	EN	MIT
NousResearch/hermes-function-calling-v1 (func_calling_singleturn)	2,500	EN	Apache-2.0

Every imported row records provenance fields such as source_dataset, source_record, source_license, and attribution. See SOURCE_MANIFEST.json for the machine-readable source manifest.

Transformations

The imported upstream text is selected and transformed into the jev-my-bro governance contract. Source labels are not presented as original governance annotations.

The provenance-aware imported families use these variants:

Variant	Purpose
base	Source-derived base request
approval_context	Verified explicit authorization context
revoked_approval	Previously granted authorization was revoked
conflict_context	Conflicting authorization/scope instructions
ood_mixed_indirect	Indirect/OOD framing with Thai-English mixed wording

Current variant counts are:

Variant	Cases
base	2,508
approval_context	1,500
revoked_approval	1,500
conflict_context	1,500
ood_mixed_indirect	1,500

Schema

Each JSONL row has this high-level shape:

{
  "id": "stable-case-id",
  "workflow": "agent_operation_governance",
  "language": "en",
  "state": {
    "request": "operation or request text",
    "domain": "domain-name",
    "difficulty": "easy|medium|hard",
    "source_dataset": "upstream-or-project-source",
    "source_record": "upstream-record-id",
    "source_license": "license-id",
    "attribution": "source attribution",
    "scenario_family": "family-id",
    "variant_type": "base|approval_context|revoked_approval|conflict_context|ood_mixed_indirect",
    "label_status": "existing-static|rule_reviewed"
  },
  "questions": {
    "action": {"type": "choice"},
    "needs_review": {"type": "noul"},
    "prohibited": {"type": "noul"},
    "risk": {"type": "score"}
  },
  "gold": {
    "action": {"label": "execute", "probabilities": {}},
    "needs_review": {"label": "false", "noul": 0.08, "probabilities": {}},
    "prohibited": {"label": "false", "noul": 0.08, "probabilities": {}},
    "risk": {"label": "0", "score": 0.2, "probabilities": {}}
  }
}

All training targets are probability distributions. For risk, score is kept consistent with the probability-weighted expected level. Score targets use a distance-aware Gaussian ordinal distribution (sigma=0.75), so neighboring risk levels receive more probability mass than distant levels. Risk measures operational impact, blast radius, and reversibility if executed; authorization and prohibition are handled by their dedicated primitives.

Action distribution

Across all splits:

Action	Cases
ask_user	4,766
execute	3,406
reject	336

This distribution is intentionally reported because the dataset is not class-balanced. Consumers should evaluate per-class behavior instead of relying only on aggregate accuracy.

Risk distribution

Risk level	Cases
0	2,623
1	68
2	3,168
3	2,313
4	336

The project-authored 1,008-case bootstrap subset was relabeled for the v4 ordinal-risk rubric so all five levels are represented. Imported rule_reviewed rows retain their semantic labels but use the same ordinal target distribution. Risk level 1 remains sparse in the expanded corpus and should be treated as a known coverage limitation until further human-reviewed curation is added.

Load from Hugging Face

from datasets import load_dataset

dataset = load_dataset("JonusNattapong/jev-my-bro-dataset")

print(dataset["train"][0])

Repository training uses the checked-out JSONL files directly:

python -m jevbro.train \
  --train data/hf_expanded/train.jsonl \
  --validation data/hf_expanded/validation.jsonl

Calibration and final evaluation use their dedicated held-out splits:

python -m jevbro.calibrate \
  --model artifacts/laya-model \
  --data data/hf_expanded/calibration.jsonl

python -m jevbro.evaluate \
  --model artifacts/laya-model \
  --data data/hf_expanded/test.jsonl

Do not fit calibration or tune hyperparameters on the test split.

Validation and quality checks

The source repository provides two checks:

python scripts/validate_dataset.py --root data/hf_expanded
python scripts/audit_hf_dataset.py --root data/hf_expanded

For the published snapshot on 2026-09-21:

• cases: 8,508
• decisions: 34,032
• scenario families: 2,508
• audit flags: 0
• normalized duplicate request leakage across splits: 0
• scenario-family leakage across splits: 0

The checks are structural and rubric-consistency checks; they do not substitute for independent human review.

Evaluation status

Previously published jev-my-bro v0.2 metrics were measured on the older 1,008-case bootstrap dataset. Those scores should not be treated as metrics for this updated 8,508-case snapshot.

A new training, calibration, and test run is required before publishing model quality claims for this dataset revision.

Licenses and attribution

This is a mixed-source dataset. There is no claim that all rows can be relicensed under one upstream license. Consumers and redistributors should respect the license and attribution metadata associated with each source.

MASSIVE

• Dataset: AmazonScience/massive
• Configuration: th-TH
• License recorded by this dataset: CC BY 4.0
• Attribution: FitzGerald et al., MASSIVE (2022); Amazon Science
• Source: https://huggingface.co/datasets/AmazonScience/massive

BANKING77

• Dataset: mteb/banking77
• License recorded by this dataset: MIT
• Attribution: Casanueva et al., BANKING77 (2020); MTEB mirror
• Source: https://huggingface.co/datasets/mteb/banking77

Hermes Function Calling

• Dataset: NousResearch/hermes-function-calling-v1
• Configuration: func_calling_singleturn
• License recorded by this dataset: Apache-2.0
• Attribution: Nous Research Hermes function-calling dataset
• Source: https://huggingface.co/datasets/NousResearch/hermes-function-calling-v1

jev-my-bro bootstrap data

• Dataset: jev-my-bro-static-v0.1
• Cases: 1,008
• License: Apache-2.0
• Project license text: https://github.com/JonusNattapong/jev-my-bro/blob/main/LICENSE

The combined dataset remains mixed-license. The project Apache-2.0 license does not replace or override MASSIVE CC BY 4.0, BANKING77 MIT, or other applicable upstream terms.

Known limitations

• rule_reviewed cases are not independently human-labeled.
• The reject action is substantially less frequent than ask_user and execute.
• The dataset is specialized for agent/tool governance and should not be interpreted as a general safety or authorization benchmark.
• Source-domain language and transformation templates can introduce shortcuts; per-source, per-language, per-variant, and per-class evaluation is recommended.
• The model consuming this data should remain advisory unless paired with explicit policy enforcement, human escalation, and audit logging.

Related project

• Source code: https://github.com/JonusNattapong/jev-my-bro
• Model: https://huggingface.co/JonusNattapong/jev-my-bro
• Dataset: https://huggingface.co/datasets/JonusNattapong/jev-my-bro-dataset