Research map / Large language models / Instruction and fine-tuning
CPQS-Tuning: A Model Self-Perception-Based Data Filtering Algorithm for Efficient Instruction Fine-Tuning
Yi Ren, Yanhui Li, Tianyi Zhang, Diandong Liu
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 changed | Where | Measured | Effect |
|---|---|---|---|
| filter to 5k CPQS-selected samples instead of full 52k | Llama2-7B on Alpaca GPT4 · Open LLM average | average accuracy: 54.37 -> 54.40 | Improved |
| filter Reasoning-DeepSeek to 10k CPQS samples instead of full 146k | Qwen2.5-7B-Instruct · Reasoning-DeepSeek average | average accuracy: 60.92 -> 65.45 | Improved |
| increase CPQS subset from 10k to 50k | Qwen2.5-7B-Instruct · Reasoning-DeepSeek average | average accuracy: 65.45 -> 65.02 | No clear effect |
| use CPQS selection instead of Superfiltering at 1k | Llama2-7B · AlpacaEval | win rate: 55.87 -> 55.98 | Improved |
| use CPQS selection on GSM8K 500 samples vs original model | Qwen2.5-7B-Instruct · GSM8K | accuracy: 76.27 -> 81.12 | Improved |
| use CPQS selection on Magicoder 1000 samples | Llama 2-7B-Chat · HumanEval avg | pass@1: 12.5 -> 14.95 | Improved |
| use CPQS selection on Magicoder 1000 samples | Qwen2.5-7B-Instruct · HumanEval avg | pass@1: 79.25 -> 77.2 | Hurt |
| use hidden states from all layers instead of only early 9 layers (ablation) | CPQS CNN selector · GSM8K | accuracy: 84.23 -> 84.91 | Improved |
| use hidden states from all layers instead of final layer only (ablation) | CPQS CNN selector · HumanEval | pass@1: 78.7 -> 80.0 | Improved |
| use CNN selector instead of MLP (ablation) | CPQS · Reasoning-DeepSeek 10k average | average accuracy: 62.38 -> 65.45 | Improved |
| use CNN selector instead of Transformer (ablation) | CPQS · Reasoning-DeepSeek 10k average | average accuracy: 63.81 -> 65.45 | Improved |
| use own-model selected data instead of swapped dataset (ablation) | Qwen2.5-7B-Instruct · GSM8K | accuracy: 83.24 -> 84.91 | Improved |
Closest papers by meaning
- LESS: Selecting Influential Data for Targeted Instruction Tuning (ICML 2024 · similarity 0.89)
- PandaLM: An Automatic Evaluation Benchmark for LLM Instruction Tuning Optimization (ICLR 2024 · similarity 0.89)
- TAIA: Large Language Models are Out-of-Distribution Data Learners (NeurIPS 2024 · similarity 0.88)
- DELIFT: Data Efficient Language model Instruction Fine-Tuning (ICLR 2025 · similarity 0.88)
- SHED: Shapley-Based Automated Dataset Refinement for Instruction Fine-Tuning (NeurIPS 2024 · similarity 0.88)
- Long Is More for Alignment: A Simple but Tough-to-Beat Baseline for Instruction Fine-Tuning (ICML 2024 · similarity 0.88)
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.
