Source-anchored instruction pairs reduce hallucination: they achieve mean Teacher Fidelity 1.4 points higher than non-anchored pairs, and models trained on anchored subsets produce 23% fewer factual hallucinations¶

Claim · All claims
Evidence · 1 study · design r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person

Subclaims¶

q2 i? Source-anchored pairs score higher on Teacher Fidelity and yield fewer factual hallucinations in held-out evaluations than non-anchored pairs. → Shwetha Singaravelu 2026

Evidence¶

Shwetha Singaravelu 2026¶

Shwetha Singaravelu, Gayathri Muruganantham, Lakshmi Rajendran, Santhosh Sivasubramani. (2026). IKS-Instruct: A 24,000-Example Multilingual Dataset for Teaching Language Models Indian Knowledge Systems. arXiv. https://arxiv.org/abs/2607.23322 doi:10.48550/arxiv.2607.23322

q2 · i? · design · r2

Data-quality analysis in Section 6.2 comparing source-anchored versus template-generated pairs without source anchoring, reporting mean TF scores of 7.8 versus 6.4 and a 23% reduction in factual hallucinations; no standardized effect size is printed.

"source-anchored pairs achieve mean TF scores 1.4 points higher than non-anchored pairs (7.8 vs. 6.4), and fine-tuned models trained on source-anchored subsets produce 23% fewer factual hallucinations in held-out evaluations"

Discussion¶

Related Claims¶

• Model-level judge quality is non-monotonic in dataset size: the 13,882-pair v1.1 yields the best model (6.39), while both expansion to ~20,000 pairs (6.33) and aggressive curation to 7,130 pairs (6.12) reduce quality — related
• LLM-assisted inductive qualitative coding carries risks of superficial themes, broad or redundant codes, and hallucinated interpretations, so LLMs should augment rather than replace human researchers — related
• Poorly grounded GenAI self-study risks fabricated content and inferior learning relative to curated resources — related
• The verification threshold trades hallucination reduction against over-refusal, with the default tau of 0.20 providing the best balance — related
• Qualitative error analysis yields five recurring error types, with over-inference most prevalent (33.5%), followed by evidence misidentification (23.4%) and hallucination (19.8%) — related
• LLMs in educational applications hallucinate in 15–25% of responses and can amplify bias against non-native language inputs — related
• Recall-oriented anchor selection achieved a 17.93% anchor rate and 94.63% non-empty OCR rate, which the authors interpret as anchors frequently capturing visible educational content that transcript chunks omit — related