🧠 GPT-2 NMC

GPT-2 NMC v8.0 - Semantic Expert Discovery (4.1x compression, auto-synonym mapping)

Loading Model...

0%
Initializing...

Loading 118.9 MB compressed + computing NMC base weights
(First load takes ~30-60 seconds as weights are reconstructed)

Ready! Enter a prompt to generate text.
Basic Controls
0.8
0.9
40
Advanced Sampling
1.2
0.5
0
0
1
(better quality, slower)
(reduce CAPS artifacts)
NMC Compression
4
1.0
1.0
Tokens
0
Speed
-- tok/s
Memory
118.9 MB
Compression
4.1×

How NMC Works

Base weights are computed on-the-fly from coordinates using Collatz dynamics. Only tiny 4-bit residuals (~99% near-zero) are loaded from storage. This achieves 4.1× compression while maintaining full GPT-2 quality.

🎯 NMC Fine-Tuning

Train the model to minimize residuals while maintaining language ability. This adjusts weights to be closer to the NMC base.

Ready to train

🔥 Train from Collatz: Initializes NMC base + small random residuals, then trains. ▶ Continue: Keep training from current state (doesn't reset).