Research map / Federated learning and privacy / Personalized and heterogeneous FL
CLUES: Collaborative Private-domain High-quality Data Selection for LLMs via Training Dynamics
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CLUES, a training-dynamics-based data quality selection method for collaborative (federated or model-merging) LoRA fine-tuning of LLMs on private silos, using a public anchor set to set a global quality threshold.
Problem
Automated data quality control for LLM fine-tuning is hard in collaborative settings where private data silos cannot share raw data and have heterogeneous quality and domains.
Approach
Score each training sample by the accumulated inner product of its per-sample loss gradients with those of a public validation/anchor set across training checkpoints, use anchor-data scores to set one global threshold shared by all clients, and then fine-tune collaboratively (TIES or Task Arithmetic merging, or FedAvg) on the filtered data.
Key results
- On medical QA with Mistral-7b in federated setting, CLUES reaches GPT-4 score 0.161 vs 0.085 for mixed-quality data (about 100.6% of Oracle's 0.160).
- On FiQA data selection, global-threshold selection achieves 97.44% precision and 97.91% accuracy, versus 95.00% accuracy for a pre-determined score and 75.00% for ratio-based selection.
- Across 0-80% low-quality ratios, CLUES selects over 99% of high-quality data (accuracy 99.31% to 99.84%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CLUES-selected data instead of mixed-quality data | Mistral-7b · Medical QA, federated | GPT-4 Scoring: 0.085 -> 0.161 | Helped |
| Use CLUES-selected data instead of mixed-quality data | Llama2-7b · Medical QA, federated | GPT-4 Scoring: 0.0952 -> 0.210 | Helped |
| Use CLUES-selected data instead of mixed-quality data | Llama2-7b · Medical QA, federated | Knowledge Avg: 0.311 -> 0.356 | Helped |
| Use CLUES-selected data instead of mixed-quality data | Mistral-7b · MMedBench, federated | score: 0.420 -> 0.435 | Helped |
| Use CLUES-selected data instead of mixed-quality data | Llama2-7b · MMedBench, federated | score: 0.440 -> 0.477 | Helped |
| Use CLUES-selected data instead of mixed-quality data | Mistral-7b · MMedBench, model merging | score: 0.515 -> 0.525 | Helped |
| Use CLUES-selected data instead of mixed-quality data | Llama2-7b · MMedBench, model merging | score: 0.485 -> 0.487 | Helped |
| Use CLUES-selected data instead of mixed-quality data | Mistral-7b · Medical QA, federated | Knowledge Avg: 0.194 -> 0.309 | Hurt |
| Use the anchor-data global threshold instead of selection by ratio (ablation) | CLUES data selection · FiQA | Accuracy: 75.00% -> 97.91% | Helped |
| Use the anchor-data global threshold instead of a pre-determined score (ablation) | CLUES data selection · FiQA | Accuracy: 95.00% -> 97.91% | Helped |
| Use TIES merging instead of Task Arithmetic (ablation) | CLUES · Financial QA | Financial QA Scores: 3.776 -> 3.855 | Helped |
| Use TIES merging instead of Task Arithmetic (ablation) | CLUES · Multilingual QA | Multilingual QA Scores: 0.470 -> 0.487 | Helped |
Surprising findings
- Linear (task-arithmetic) merging with CLUES data selection did not reach the performance of low-quality data with TIES merging, showing merging method matters a lot.
Limitations
- Assumes all local models share the same architecture (LoRA with same rank), limiting heterogeneous-architecture settings.
- Evaluation relies on GPT-4 scoring on small test sets (about 200 samples) and synthetic low-quality pollution.
Experimental setup
Datasets: PMC-LLama Medalpaca-flashcards MMedBench FiQA MMLU clinical topics MedMCQA PubMedQA USMLE
Benchmarks: MMedBench FiQA MMLU clinical topics MedMCQA PubMedQA USMLE
Metrics: GPT-4 score knowledge average accuracy precision recall F1 accuracy
Baselines: Oracle Perplexity score IFD DataInf
Builds on: DataInf IFD TIES merging Task Arithmetic LoRA
Closest papers by meaning
- Improving Data Efficiency via Curating LLM-Driven Rating Systems (ICLR 2025 · similarity 0.87)
- Fixing It in Post: A Comparative Study of LLM Post-Training Data Quality and Model Performance (NeurIPS 2025 · similarity 0.87)
- DataMan: Data Manager for Pre-training Large Language Models (ICLR 2025 · similarity 0.87)
- Data Mixing Optimization for Supervised Fine-Tuning of Large Language Models (ICML 2025 · similarity 0.87)
- FlowerTune: A Cross-Domain Benchmark for Federated Fine-Tuning of Large Language Models (NeurIPS 2025 · similarity 0.86)
- QuRating: Selecting High-Quality Data for Training Language Models (ICML 2024 · similarity 0.86)
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.