Polydim-Einsof's picture
Feat: Einsof-II LoRA Adapter trained on Kaggle Dual T4 (Loss: 0.5710, Stiefel Defect: 2.77e-16)
95fa1f8 verified
|
Raw History Blame Contribute Delete
813 Bytes
---
license: apache-2.0
base_model: Qwen/Qwen2.5-1.5B-Instruct
tags:
- pmtp
- stiefel
- einsof-ii
- mathesis
- radical-honesty
---
# Einsof-II LoRA Adapter (Kaggle Dual Tesla T4 Silicon Hardened)
Entrenado sobre el dataset unificado `Polydim-Einsof/einsof2-training-datasets` (1,812 muestras).
Incorpora geometría de Stiefel $St(k, d)$, Protocolo Fantasma (Ghost Protocol) y arquitectura de Dos Módulos Interconectados (Atelier Co-Design).
## Métricas de Silicio Físico (Dual NVIDIA Tesla T4)
- **Base Model:** `Qwen/Qwen2.5-1.5B-Instruct`
- **Defecto Isométrico Stiefel ($S^{2047} \to St(1024, 2048)$):** `2.7743304968121066e-16`
- **Tier de Precisión:** `TIER_MACHINE_EPS_16` (Piso mantisa IEEE 754)
- **Loss de Entrenamiento:** `0.5710`
- **Exit Code:** `0` (Radical Honesty - Regla Constitucional 1)