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¶
Claim · All claims
Evidence · 1 study · designr3·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¶
Reading the evidence codes
q is the design tier, i the impact, n the sample, and the kind with its r
the rigour judged within that kind: what each code means.
This claim: no standardized effect is recorded in observations/ for it yet, so no benchmark reading is possible.
q2 i? Neither raw volume expansion nor aggressive curation improved model-level judge scores relative to the moderately sized v1.1 dataset. → 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 · r3
Version-level comparison in Section 5.3 (Table 6) of three LoRA models trained on successive IKS-Instruct versions, scored by the same five-judge panel. Median scores are 6.39 (v1.1), 6.33 (v1.7), and 6.12 (v2.1); no effect size is printed.
"The moderately sized 13,882-pair v1.1 produced the best model (6.39); expanding to the 20,000-pair v1.7 slightly reduced quality (6.33), consistent with the noisy auto-generated Vedic-mathematics pairs added in that step; and aggressively curating down to the 7,130-pair v2.1 reduced quality further at the model level (6.12), even though curation improved several dataset-level quality metrics."
Discussion¶
Related Claims¶
- IKS-Instruct fine-tuning is necessary for IKS pedagogical capability: a compact 7B fine-tune reaches a median judge score of 6.39, within 0.15 of a general-purpose reference, while the base model scores near zero on IKS-specific dimensions — related
- 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 — related
- The highest decomposition level (five segments) performed slightly worse than the three-segment approach, suggesting non-monotonic benefits of decomposition — a broader claim this one bears on
Page metadata
| Field | Value |
|---|---|
type |
claim |
id |
non-monotonic-quality-volume-iks-instruct |
title |
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 |
description |
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 |
status |
review |
evidence_strength |
weak |
newest_evidence |
2026 |
recheck_by |
2029 |
generated |
by process:wiki-ingest on 2026-10-11 |
trust |
unverified — no one has confirmed this page's content |
sources |
1 — shwetha-singaravelu-2026 |
Frontmatter is the page's machine-readable record — see the Schema & Guide.