jev-my-bro-th960

Jev laya-th960

Jev laya-th960 is a Thai typed decision model for agent and tool governance. Given an operational request, it returns four signals: an action choice, whether human review is needed, whether the operation is prohibited, and an ordinal operational-risk score from 0 to 4.

This is a research and engineering checkpoint. It is an advisory signal and must not replace explicit authorization, repository policy, human review, audit logs, or an independent safety layer.

Model details

• Base model: convaiinnovations/laya-multilingual
• Foundation: Laya 0.3.4
• Language: Thai (th)
• Decision primitives: choice, noul, and ordinal score
• Training: RLCD, soft-target cross-entropy, ordinal cumulative loss, and RPS
• Selected checkpoint: epoch 5, selected from validation metrics
• Calibration: one temperature per primitive plus score thresholds
• Decoder: threshold, selected from the calibration split only
• Repository: JonusNattapong/jev-my-bro-th960

Evaluation

The final locked evaluation used data/th_curated_960/test.jsonl with 100 cases and 400 typed decisions. Each risk level has 20 cases. The test split was not used for training, temperature fitting, threshold fitting, or decoder selection.

Accuracy by primitive

Score recall by risk level

Risk 1 and risk 3 are the weakest per-level slices. The ECE of 0.1391 also shows that confidence is not perfectly calibrated; consumers should use the probabilities for escalation and monitoring, not as guarantees.

Usage

The repository is private unless the owner changes its visibility. Authenticate with Hugging Face without putting a token in source code:

from huggingface_hub import notebook_login
notebook_login()

Then load the model with the Jev question schema:

import json
import torch
from laya import Agent
from jevbro.questions import default_questions

agent = Agent(
    "JonusNattapong/jev-my-bro-th960",
    device="cuda" if torch.cuda.is_available() else "cpu",
)

result = agent.predict(
    {
        "request": "เปลี่ยนค่า firewall ของ production ให้เปิดจากอินเทอร์เน็ต",
        "domain": "network",
    },
    default_questions("th"),
)

print(json.dumps(result["answers"], ensure_ascii=False, indent=2))

Keep the policy gate outside the model. A prohibited signal has priority over the raw action choice:

answers = result["answers"]
if answers["prohibited"]["noul"] >= 0.5 or answers["action"]["choice"] == "reject":
    decision = "reject"
elif answers["needs_review"]["noul"] >= 0.5 or answers["action"]["choice"] == "ask_user":
    decision = "human_review"
else:
    decision = "execute"

Qualitative smoke example

For the request เปลี่ยนค่า firewall ของ production ให้เปิดจากอินเทอร์เน็ต, the raw model returned action=ask_user, needs_review=0.809, prohibited=0.605, and risk_score=2.877. The external policy gate therefore resolved the request to reject. This illustrates why the model output must be combined with an explicit policy gate.

Intended use and limitations

Intended uses include Thai-language governance experiments, routing requests to human review, risk triage, and testing typed decision interfaces. The model is not intended to autonomously authorize production changes, handle secrets, make legal or compliance decisions, or serve as the only control for destructive operations.

The evaluation set is manually curated development data. It does not prove performance on unseen organizations, languages, domains, or production traffic. Validate the model with reviewed operational data and real outcome feedback before using it in a consequential workflow.

Dataset and reproducibility

The matching dataset is JonusNattapong/jev-my-bro-dataset-th960. The repository contains the source rows, split builder, T4 configuration, and locked calibration/evaluation workflow:

python scripts/build_th_curated_960.py
python scripts/validate_dataset.py --root data/th_curated_960
python -m jevbro.train --config configs/colab-th960.yaml
python -m jevbro.calibrate \
  --model artifacts/laya-th960 \
  --data data/th_curated_960/calibration.jsonl \
  --report artifacts/laya-th960/calibration-report.json
python -m jevbro.evaluate \
  --model artifacts/laya-th960 \
  --data data/th_curated_960/test.jsonl \
  --device cuda \
  --lock-decoder \
  --report artifacts/laya-th960/test-report.json

License

The project code and project-authored data are licensed under Apache-2.0. Laya is also Apache-2.0. Review the repository's docs/THIRD_PARTY.md before redistributing derived artifacts.