VietEduQwen achieves 87.02% accuracy on the 2025 Vietnamese National High School Examination, a 6.10-percentage-point gain over the base Qwen3-8B model¶

Claim · All claims
Evidence · 1 study · design r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? Fine-tuning Qwen3-8B with SFT and DPO raises examination accuracy from 80.92% to 87.02% across four epochs on 3,119 exam questions. → Thang Doan Viet 2026

Evidence¶

Thang Doan Viet 2026¶

Thang Doan Viet, Anh Nguyen Hoang, Tinh Luong Son, Anh Hoang Thi Ngoc, Huyen Giang Thi Thu, Tai Le Quy. (2026). ConnectED: A Curriculum-Aligned AI System for Vietnamese Instructional Lesson Planning and Student Learning. arXiv. https://arxiv.org/abs/2607.28647 doi:10.48550/arxiv.2607.28647

q2 · i? · design · r2

Benchmark evaluation of VietEduQwen on 3,119 questions from the 2025 Vietnamese National High School Examination covering eight subjects. Accuracy rose from "80.92% for the base Qwen3-8B model to 87.02% after Epoch 4"; the paper also reports it outperforming Gemma-2-9B and comparable to Gemini 3 Flash.

"overall examination accuracy increases from 80.92% for the base Qwen3-8B model to 87.02% after Epoch 4, corresponding to a total gain of +6.10%."

Discussion¶

Related Claims¶

• Removing the DPO objective reduces examination accuracy by 1.11 percentage points and degrades annotated pedagogical quality — related
• VietEduQwen improves over its base model across all eight exam subjects, with the largest gains in STEM disciplines, and its residual errors concentrate in multi-step computation and cross-concept integration — a narrower finding that bears on this claim
• CDPK performance scales with model size, with a sharp Pareto-frontier drop-off below around 8B parameters — related