Research map / Large language models / Instruction and fine-tuning
Pre-training LLM without Learning Rate Decay Enhances Supervised Fine-Tuning
Kazuki Yano, Shun Kiyono, Sosuke Kobayashi, Sho Takase, Jun Suzuki
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 |
|---|---|---|---|
| remove LR decay in pre-training (WSO instead of best decay scheduler) | Llama-3 1B · after SFT | SFT Task Avg: +0.3 vs best decay | Improved |
| remove LR decay in pre-training (WSO) | Llama-3 8B · after SFT | SFT Task Avg: +1.1 vs best decay | Improved |
| remove LR decay in pre-training (WSO) | Llama-3 1B and 8B · after pre-training | PT valid loss and PT Task Avg: loss +0.071 (1B), +0.127 (8B); task avg -1.7 (1B), -0.8 (8B) | Hurt |
| decay LR to zero instead of 10% (alpha 0.1 to 0.0) in WSD (ablation) | Llama-3 1B · after SFT | SFT Task Avg: +0.0 -> -1.0 | Hurt |
| use constant LR in mid-training instead of linear decay (alpha_mid 1.0 vs 0.0) (ablation) | WSD pre 0.1, 1B · after SFT | SFT Task Avg: -1.3 -> -0.5 | Improved |
| use constant LR through pre- and mid-training (WSO) in three-stage pipeline | Llama-3 8B · after SFT | SFT Task Avg: +1.1 vs best decay | Improved |
| use WSO in over-training with 2T tokens | Llama-3 1B · after SFT | SFT Task Avg: +0.7 | Improved |
| use WSO in over-training with mid-training | Llama-3 1B · after SFT | SFT Task Avg: +1.4 | Improved |
| remove LR decay (WSO) to reduce sharpness | Llama-3 1B · sharpness of pre-trained model | Hessian trace: decay models 2-3x higher sharpness | Improved |
| higher sharpness of pre-trained model (ablation) | Llama-3 1B · SFT performance | SFT score: Pearson r = -0.709 | Hurt |
Closest papers by meaning
- Through the River: Understanding the Benefit of Schedule-Free Methods for Language Model Training (NeurIPS 2025 · similarity 0.91)
- A Multi-Power Law for Loss Curve Prediction Across Learning Rate Schedules (ICLR 2025 · similarity 0.90)
- Estimating and Orthogonalizing Unknown Pre-training Gradients for Continual Fine-tuning of Large Language Models (NeurIPS 2026 · similarity 0.90)
- Revisiting Zeroth-Order Optimization for Memory-Efficient LLM Fine-Tuning: A Benchmark (ICML 2024 · similarity 0.89)
- Benchmarking Optimizers for Large Language Model Pretraining (NeurIPS 2026 · similarity 0.89)
- Two-stage LLM Fine-tuning with Less Specialization and More Generalization (ICLR 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.
