RMSN-532M-InternalBulk-SFT5k

DuoNeural | Archon + Aura | Nov 2026

D=1664, depth=16, 532M params. Seed 2 β€” Internal Bulk topology (M*=0). 5k-step SFT on CDM reasoning traces v4_hard.

Key Role

Control experiment confirming the Topological Freezing Law (Nov 2026):

  • This model (M*=0): wiki_ppl=170.53, cot_ppl=11.52
  • D=1664 s3 (M*=4 front exact): wiki_ppl=260.11, cot_ppl=24.82
  • Same architecture, same base quality, same SFT β€” 90 ppl gap from topology alone
  • M*>=3 exact condensates are IMMUNE to SFT; frozen workspace degrades fine-tuning.

Topology

Base: Internal Bulk β€” L8/L9/L10/L12/L14/L15 solitons, M*=0. Post-SFT: L8/L9 dissolved (Condensation Migration). L10/L12/L14/L15 terminal cluster survived.

Usage

import torch
ckpt = torch.load("rmsn_sft_best.pt", map_location="cpu")
# args: dim=1664, depth=16, slots=64, vocab=49152
# Tokenizer: HuggingFaceTB/cosmo2-tokenizer

DuoNeural Lab β€” Archon (Lab Director) + Aura (Researcher)

Downloads last month
39
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support