Two open checkpoints show a full JEPA-pretrain-to-GRPO RL recipe on small LLMs
burny_tech · x · 2026-10-27
Author released a two-stage JEPA→GRPO training pipeline: a Qwen2.5-Math-1.5B checkpoint JEPA-pretrained (LLM-JEPA cosine alignment + SIGReg, official recipe) on 32k verified math CoT traces, converging at cos≥0.90 in 400 steps and serving as init for stage-2 GRPO RL; plus openmed-community/granite-4.0-micro-OpenMed, a 3B Granite 4.0 Micro tuned via JEPA-LLM SFT (medmcqa-hard + personas) then GRPO and an Arcee Fusion merge, Apache-2.0 licensed. Together they lay out a reproducible path combining JEPA representation pretraining with RLVR on small models.
More from Models
- Human-written essay flagged as Mixed by AI detector Pangram in blogger's test — Aizkmusic · 2026-10-27
- AI community buzzes: Artificial Analysis hints a major release is imminent — rand_longevity · 2026-10-27
- GPT 6.1 Sol outperforms GPT 6.1 Terra at the same cost in nearly every benchmark — Lucky_Creme_5208 · 2026-10-27
- Cryptographer: No significant cryptanalysis result from frontier AI labs since late July — matthew_d_green · 2026-10-27
- Kimi Team proposes Attention Residuals, replacing fixed residual accumulation with learned attention — burny_tech · 2026-10-27
- OpenAI claims Exponential Time Hypothesis is false, solving 3SAT in subexponential time — fortnow · 2026-10-27