SymbioGPT-Gemma-Fused-v2

Cross-species LoRA projection from Gemma-270M into SymbioGPT-10M with blend alpha = 0.5 (stronger injection). This is v2 of SymbioGPT-Gemma-Fused which used alpha = 0.3.

What Changed vs v1

Property	v1	v2
Blend alpha	0.3	0.5
Delta/weight ratio	1.4% - 4.0%	2.3% - 6.6%
PCA calibration	Same	Same (99.0% avg variance)

Everything else is identical: same source LoRA (rank 44, alpha 88, philosophy domain), same PCA projection matrices, same GQA-to-MHA head mapping.

Architecture

See v1 model card for full details on:

• Architecture mapping (Gemma GQA → SymbioGPT MHA)
• PCA projection method (640 → 320 dims, 99.0% variance)
• Layer mapping (18 → 8 proportional grouping)
• Head selection (top-5 Q-heads by delta L2 norm)

Usage

import torch
checkpoint = torch.load("symbio_gemma_fused.pt", map_location="cpu")

Load into a SymbioGPT model instance from symbiogenesis.

Links

• v1 (alpha=0.3): LisaMegaWatts/SymbioGPT-Gemma-Fused
• Source LoRA: LisaMegaWatts/SymbioSLM-ouroboros-lora-20260301
• Base Gemma: LisaMegaWatts/Ouroboros-1MContext-Gemma-270m
• Framework: symbiogenesis