Portfolio
2023-2024Technique

UNA — Uniform Neural Alignment

Yet-Unpublished LLM SFT/RLHF Technique

8 Public Releases, Multiple #1 Positions

UNA is Uniform Neural Alignment — a transformers architecture change introducing an auxiliary loss, applied as a patch to HuggingFace Transformers models. Operates during SFT and RLHF training. Applicable to attention layers, MLP layers, or both. Memory intensive but compatible with LoRA. Training data does not need to be novel, but must not have been previously overfitted. Applied across Mistral, Intel, Yi/Smaug, Qwen2.5, LLaMA 1 & 2, Pythia, and Luxa architectures.

Client
Independent — Juanako.AI
Role
Sole author · Method, training, releases
Duration
2023 & 2024
Team
Solo
Outcomes
18
Public Releases
4+
Base Architectures
1.5B to 34B
Model Sizes
Multiple
#1 Positions
Highlights
  • Consistent positive delta over base models
  • Multiple #1 leaderboard positions
  • Applicable to different network layers
Selected Releases & Leaderboard Positions
9 of 18 total releases
DateModelSizeBase1.5B3B7B34B70B+
28-Nov-2023Juanako 7B UNA7BMistral··🏆··
02-Dec-2023Cybertron v17BMistral··🏆··
05-Dec-2023Cybertron v27BMistral··🏆🏆🏆
09-Dec-2023Xaberius 34B v1beta34BYi···🏆🏆
11-Jan-2024UNA-TheBeagle v17BMistral··🏆🏆🏆
04-Feb-2024UNA-SimpleSmaug 34B34BYi/Smaug···🏆·
30-Oct-2024Cybertron v4 MGS7BQwen2.5··🏆··
07-Nov-2024MiniClaus UNAMGS1.5BQwen2.5🏆····
21-Nov-2024Cybertron v4 UNAMGS7BQwen2.5··🏆··
🏆 marks the size tier reached on the leaderboard
transformersdeepspeedaccelerateaxolotltorchwandbsftrhlfdistributed-training