Portfolio
2024-2025Technique

MGS — MultiGumbelSampling

Yet-Unpublished LLM Regularization Technique

Compatible with UNA for Additive Gains

MGS is MultiGumbelSampling — a regularization technique introducing Gumbel-sampled noise across signal paths during SFT/RLHF training. Combinable with UNA (UNAMGS releases) for additive performance gains.

Client
Independent — Juanako.AI
Role
Sole author
Duration
2024-2025
Team
Solo
Outcomes
5
Public Releases
4
UNAMGS Releases
1.5B to 7B
Model Sizes
Multiple
#1 Positions
Highlights
  • Compatible with UNA — UNAMGS combines both
  • Operates on different network paths than UNA
  • First public release: Oct 2024
Selected Releases & Leaderboard Positions
DateModelSizeBase1.5B3B7B34B70B+
30-Oct-2024Cybertron v4 MGS7BQwen2.5··🏆··
07-Nov-2024MiniClaus UNAMGS1.5BQwen2.5🏆····
21-Nov-2024Cybertron v4 UNAMGS7BQwen2.5··🏆··
04-Nov-2024Pancho v1 3B UNAMGS3BQwen2.5·🏆···
03-Feb-2025MiniClaus UNAMGS GRPO1.5BQwen2.5🏆····
🏆 marks the size tier reached on the leaderboard
transformersdeepspeedaccelerateaxolotltorchwandbsftrhlfdistributed-training