AI Paper Atlas

Research map / Large language models / Instruction and fine-tuning

Cure-SFT: Diagnostic-Guided Data Curation for Instruction Tuning

Yuankang Fu, Xinrong Gong, Chen Gong, Tong Zhang, Kaixiang Yang

ICML 2026 · Large language models › Instruction and fine-tuning · approach: Fine-tuning for reasoning data

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add stratified semantic-geometric sampling (Stage I) (ablation)Cure-SFT · Alpaca 10%average winning score: 1.72 -> 1.78Improved
add diagnostic triage (Stage II) and remediation on top of sampling (ablation)Cure-SFT · Alpaca 10%average winning score: 1.05 -> 1.19Improved
add targeted remediation of fixable samples (Stage III) (ablation)Cure-SFT · Alpaca 10%average winning score: 1.19 -> 1.78Improved
random sampling instead of semantic-geometric sampling (ablation)Cure-SFT · Alpaca 10%, Vicunawinning score: 1.93 vs 1.91No clear effect
targeted remediation instead of unified rewritingCure-SFT vs Rewrite · Alpaca 10%, AlpacaEval 2.0 LC winLC win rate: 12.98 -> 11.45Hurt
targeted remediation instead of unified rewritingCure-SFT vs Rewrite · Alpaca 10%average winning score: 1.54 -> 1.78Improved
use selection-based methods (IFD, Deita) on higher-quality dataIFD, Deita · WizardLM 10%average winning score: IFD 0.67, Deita 0.62 vs Random 0.86Hurt
increase data budget from 10% to 15%Cure-SFT · Alpaca, WizardLM testwinning score: 1.85 -> 1.78No clear effect
use 1% data budgetCure-SFT · Alpaca, WizardLM testwinning score: 1.60 at 1%Improved
use smaller teacher Qwen2.5-14B instead of 72B (ablation)Cure-SFT · Alpaca 10%winning score: 1.48 to 1.90 vs 72B slightly higherNo clear effect

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy