bluemorpholimited/Morpho-72B-MoE-SFT
<p align="center"> <img src="https://img.shields.io/badge/adapter-singleexpertpreview-ff2d78?style=for-the-badge" /> <img src="https://img.shields.io/badge/rank-64-b967ff?style=for-the-badge" /> <img src="https://img.shields.io/badge/status-superseded-8b9bb4?style=for-the-badge" /> <img src="https://img.shields.io/badge/license-Apache_2.0-8b9bb4?style=for-the-badge" /> </p>
A relic from the early runs. This repo holds a single rank-64 preview adapter from the first SFT experiments — kept for provenance, not part of the shipping model.
Where things actually live
The production series is rank-128, FP32, 1,120 tensors per expert, each header-verified at exactly 6,737,255,616 bytes. This preview (rank-64, 3.37 GB) predates that standard — use the Vault for anything real.
Usage (preview only)
Load with PEFT over the same 72B dense backbone used across this project — any standard PeftModel.from_pretrained(base, ...) flow works. This adapter is rank-64 and unconverged by current standards; prefer the Vault.
⚠️ Research artifact. For the converged per-expert series, see the Checkpoint Vault.
License
Apache 2.0.
<p align="center"><i>KEPT FOR PROVENANCE — SUPERSEDED BY THE VAULT</i></p>
