Under a matched 2,300-sample SFT budget, K12-Train-Text outperforms eight mainstream instruction-tuning corpora on GaokaoBench and EduEval across two base models¶

Claim · All claims
Evidence · 1 study (2 entries) · causal r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person

Subclaims¶

q2 i? On Qwen3-4B-Base, K12-Train-Text achieves the highest GaokaoBench total (1009.96), improving over the strongest baseline DataFlow (985.91) by +24.1; on Llama3.1-8B-Base it again obtains the highest total (625.49), outperforming WizardLM (593.08) by +32.4. → Hao Liang 2026 q2 i? On EduEval, K12-Train-Text achieves the best average score on both backbones (66.76 on Qwen3-4B-Base; 40.90 on Llama3.1-8B-Base) under the same 2,300-sample budget. → Hao Liang 2026 (2)

Evidence¶

Hao Liang 2026¶

Hao Liang, Qihan Lin, Mingrui Chen, Hengyi Feng, Zhaoyang Han, Xiaochen Ma, Zhen Hao Wong, Meiyi Qiang, Linzhuang Sun, Wentao Zhang. (2026). K12-KGraph: A Curriculum-Aligned Knowledge Graph for Benchmarking and Training Educational LLMs. arXiv preprint. https://arxiv.org/abs/2605.09635 doi:10.48550/arxiv.2605.09635

q2 · i? · causal · r2

Full-parameter SFT of Qwen3-4B-Base and Llama3.1-8B-Base on K12-Train-Text and eight general instruction-tuning datasets, all matched to approximately 2,300 samples with identical hyperparameters, evaluated on GaokaoBench (Table 4). The article reports the highest totals for K12-Train-Text on both backbones.

"On Qwen3-4B-Base, K12-Train-Text achieves the highest overall total (1009.96), improving over the strongest baseline DataFlow (985.91) by +24.1, while also reaching the best objective score rate (81.8%) and subjective score rate (89.5%)."

Hao Liang 2026 (2)¶

Hao Liang, Qihan Lin, Mingrui Chen, Hengyi Feng, Zhaoyang Han, Xiaochen Ma, Zhen Hao Wong, Meiyi Qiang, Linzhuang Sun, Wentao Zhang. (2026). K12-KGraph: A Curriculum-Aligned Knowledge Graph for Benchmarking and Training Educational LLMs. arXiv preprint. https://arxiv.org/abs/2605.09635

q2 · i? · causal · r2

Evaluation of the same SFT models on 18 EduEval sub-tasks across five capability dimensions (Table 5). The article reports K12-Train-Text reaching the best average on both backbones, though "the overall gains are smaller than on GaokaoBench, they remain consistent across both base models."

"On Qwen3-4B-Base, K12-Train-Text achieves the best average score (66.76) among all 2,300-sample SFT configurations, surpassing the strongest baseline WizardLM (66.70) while delivering the strongest Ethics results overall."

Discussion¶

Related Claims¶

• SFT on K12-Train improves out-of-domain subjects (Chinese, Humanities Math), suggesting acquisition of a transferable structurally-grounded answer style — related
• K12-Train-Full achieves the best overall performance on three multimodal educational benchmarks, and textual and visual supervision are complementary — related
• Education-oriented instruction tuning consistently improves curriculum grounding, K-12 problem solving, and pedagogical tutoring at every evaluated model scale — related