Everything I Learned Training Frontier Small Models
summary
Claims that a 350M model is trained with the LFM2.5 recipe on 28T tokens and improves benchmarks matter: showed better GPQA, IFbench, case‑report extraction and tool use versus prior LFM2. The concrete fixes matter: on‑policy preference alignment with temperature rollouts plus an LLM jury and RL with verifiable rewards and n‑gram penalties reduces doom loops from ~15% after pretrain to near zero after RL. Also note gated short convolutions for latency and agentic reinforcement learning to overcome memory limits.
AI-written summary · about
chapters (11)
- 0:00Start
- 0:14Introduction to frontier small models at Liquid AI
- 1:02Characteristics: memory-bound, task-specific, latency-sensitive
- 2:20Architecture: why large embedding layers are inefficient
- 4:01LFM2 architecture: using gated short convolutions for speed
- 6:09LFM 2.5 recipe: 28T tokens and post-training stages
- 8:34Post-training: SFT, preference alignment, and RL best practices
- 10:43Identifying "doom loops" in reasoning models
- 11:34Solutions: mitigating loops via preference alignment and RL
- 15:29Future focus: using agentic tools to overcome memory limits
- 17:58Q&A: real-world applications for small vs. large models
related talks
Stop Paying for Frontier Models
30mUnderstanding the end-to-end LLM training and inference pipeline
01h 38mTraining Large Language Models: Practices and Research Questions
42mPost-Training Small Language Models: the adventures of a practitioner
27mBuilding Mini Minds Our Adventures in Teaching Small LLMs to Reason
28mFine-Tuning Small Language Models with Python
31mBuilding High-Quality & Trustworthy FMware
39mThe Future is Fine-Tuned: Deploying Task-specific LLMs