{"type":"post","id":"pre-training-without-learning-rate-decay-4936bf","slug":"pre-training-without-learning-rate-decay-4936bf","title":"Pre-Training Without Learning Rate Decay","description":"Sebastian Raschka points to a possible change in large language model pretraining recipes: constant learning rates paired with checkpoint or weight averaging, rather than a final learning-rate decay phase.","retrieval_nugget":"Sebastian Raschka points to a possible change in large language model pretraining recipes: constant learning rates paired with checkpoint or weight averaging, rather than a final learning-rate decay phase. His examples suggest this approach is being used by Kolibri-1 and DatologyAI models, but do not establish that it outperforms established schedules.","published_at":"2026-11-18","updated_at":"2026-11-18","record_date":"2026-11-18","date_kind":"published_at","topics":["models"],"entities":["Sebastian Raschka","Kolibri-1","DatologyAI","Llama 3","OLMo 3","DeepSeek V4","Kimi K3"],"related_patterns":[],"novelty_label":"new_fact","what_changed":"Sebastian Raschka outlines pretraining recipes that replace learning-rate decay with a constant rate and checkpoint averaging, comparing approaches used across several LLMs.","prior_refs":[],"source_url":"https://sebastianraschka.com/blog/2026/llm-pretraining-without-learning-rate-decay.html","source_urls":["https://sebastianraschka.com/blog/2026/llm-pretraining-without-learning-rate-decay.html"],"schema_version":"newruntime-agent-readable-v0.2","stable_id":"post:pre-training-without-learning-rate-decay-4936bf","status":"published","visuals":[],"editor":{"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]},"routes":{"html":"https://newruntime.com/posts/pre-training-without-learning-rate-decay-4936bf/","markdown":"https://newruntime.com/posts/pre-training-without-learning-rate-decay-4936bf.md","json":"https://newruntime.com/posts/pre-training-without-learning-rate-decay-4936bf.json"},"markdown_url":"https://newruntime.com/posts/pre-training-without-learning-rate-decay-4936bf.md","next_reads":[{"type":"topic","path":"/topics/models/","url":"https://newruntime.com/topics/models/","title":"Models","reason":"Explore the models topic hub.","media_type":"text/html","markdown_url":"https://newruntime.com/topics/models.md","json_url":"https://newruntime.com/topics/models.json"},{"type":"collection","path":"/posts/","url":"https://newruntime.com/posts/","title":"Field Notes","reason":"Open the latest editorial analysis.","media_type":"text/html","json_url":"https://newruntime.com/posts.json"},{"type":"collection","path":"/sources/","url":"https://newruntime.com/sources/","title":"Sources","reason":"See the original sources this archive cites.","media_type":"text/html"}]}
