RTH-LM Math 25B - Math v1 Soul
RTH-LM Math 25B is a math-specialized Soul checkpoint for the RTH-LM/ZetaGrid architecture.
This repository contains the Math v1 smoke Soul trained with Fractal Resonant Optimization (FRO) from the Text v2 Soul base.
Artifact
souls/math_v1/MATH_V1_SMOKE.pt
Training Summary
- Base Soul:
TEXT_V2_BEST_0p9111.pt - Dataset:
math_v1_1gb.bin - Mode:
math_v1 - Rank: 512
- Optimizer: FRO
- Best observed smoke loss: 1.3773 at step 930
- Final smoke step: 1000
Notes
This is a smoke checkpoint, not a final release checkpoint. It is intended for validation, continuation training, and controlled evaluation.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support