Teacher Core
Base conversational examples for principle-level classification and explanation.
Back to projects
Rukun Ready AI / Entermind
Rukun Ready AI
The engineering story behind Rukun-32B-V: a 33B-parameter model trained for Rukun Ready AI and published through RukunNegara.ai to validate content against Malaysia's five Rukun Negara principles.
66,516
Training Records
1,353
Validation Records
8,284
Training Steps
0.2147
Eval Loss
Output Contract
Built for deterministic validation, not loose chat.
{
"principles": ["status", "severity"],
"severityBand": "safe | caution | violation",
"violationCount": 0,
"severityScore": 0.00,
"isProblematic": false,
"rewrite": "policy aligned alternative"
}principle-level statusseverity scoresaggregate severity bandnatural-language explanationviolationCountseverityScoreisProblematicpolicy-aligned rewrite
Published Under
Rukun Ready AI is the public company-facing product layer.
The research paper identifies Entermind AI as the publishing organization and points the public project site to RukunNegara.ai. This page separates your training and deployment work from the official product surface.
Rukun Negara Coverage
1
Belief in God
2
Loyalty to King and Country
3
Upholding the Constitution
4
Rule of Law
5
Good Behaviour and Morality
Training Pipeline
Corpus assembly focused on coverage, structure, and multilingual reality.
Rewrite Boost
Policy-aligned rewrite examples for non-compliant inputs.
Principle Boost
Coverage reinforcement across all five Rukun Negara principles.
Format Guard
Schema obedience records to keep output strict and machine-readable.
Records were assembled through stratification, deduplication, normalization, and audits, covering Bahasa Malaysia, English, Mandarin, Tamil, and code-switched Bahasa Rojak.
Fine-Tuning and Deployment
Base
Qwen2.5-32B-Instruct
Tuning
LoRA r=32 alpha=64
Compute
2x B200 GPUs
Serving
vLLM on RunPod
The model was fine-tuned with completion-only masking and deployed through vLLM on RunPod with deterministic decoding for moderation-grade consistency.
Endpoint Style
/v1/chat/completions
Serving Layer
vLLM / RunPod
Latency Window
~0.84s avg
Decode Policy
temperature 0
Held-Out Benchmark
88.0%
Accuracy
83.3%
Precision
90.9%
Recall
86.96%
F1
The held-out labeled benchmark contained 50 examples, with reported violating-class recall at 90.9% and F1 at 86.96%.