Research map / Large language models / Instruction and fine-tuning
LAMDAS: LLM as an Implicit Classifier for Domain-specific Data Selection
Jian Wu, Hang Yu, Bingchang Liu, Yang Wenjie, Peng Di, Jianguo Li, Yue Zhang
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 |
|---|---|---|---|
| select CPT data with LAMDAS instead of training on the full 100B-token data | Qwen2.5-7B · CPT code benchmarks | avg pass@1 (%): 28.7 -> 35.8 (+24.7%) | Improved |
| select CPT data with LAMDAS instead of DsDm (second best) | Qwen2.5-7B · CPT code benchmarks | avg pass@1 (%): 34.5 -> 35.8 | Improved |
| select CPT data with LAMDAS instead of random 15B tokens | Qwen2.5-0.5B · CPT code benchmarks | avg pass@1 (%): 15.4 -> 23.5 | Improved |
| use full CPT data instead of random 15B subset | Qwen2.5-7B · CPT code benchmarks | avg pass@1 (%): 29.6 -> 28.7 | Hurt |
| select SFT data with LAMDAS instead of full 1800K data | Qwen2.5-32B-Instruct · SFT code benchmarks | avg pass@1 (%): 57.5 -> 69.0 (+20.0%) | Improved |
| select SFT data with LAMDAS instead of LESS | Qwen2.5-32B-Instruct · SFT code benchmarks | avg pass@1 (%): 66.2 -> 69.0 | Improved |
| select SFT data with LAMDAS instead of DSIR (n-gram similarity) | Qwen2.5-7B-Instruct · SFT code benchmarks | avg pass@1 (%): 37.6 -> 44.6 | Improved |
| use random 750K SFT subset instead of base model | Qwen2.5-7B-Instruct · SFT code benchmarks | avg pass@1 (%): 29.8 -> 35.0 vs LAMDAS 44.6; relative to Full 37.3, -0.6% | Hurt |
| use selection threshold tau=1 (ablation) | LAMDAS · data selection sensitivity | downstream performance | Improved |
| use prefix length 30 (ablation) | LAMDAS · data selection sensitivity | downstream performance | Improved |
| use small classifier LLM (0.5B) for selection (ablation) | LAMDAS · data selection efficiency | selection time: 5s/100 samples vs DsDm 7s/100, LESS 12.5s/100 | Improved |
Closest papers by meaning
- BLADE: Scalable Bi-level Adaptive Data Selection for LLM Training (NeurIPS 2026 · similarity 0.89)
- Get more for less: Principled Data Selection for Warming Up Fine-Tuning in LLMs (ICLR 2024 · similarity 0.86)
- LLM Data Selection and Utilization via Dynamic Bi-level Optimization (ICML 2025 · similarity 0.86)
- Enhancing LLMs via High-Knowledge Data Selection (AAAI 2025 · similarity 0.86)
- Data Selection via Optimal Control for Language Models (ICLR 2025 · similarity 0.86)
- Enhancing Multilingual LLM Pretraining with Model-Based Data Selection (NeurIPS 2025 · 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.
