MemSFT v1.0.0

MemSFT v1.0.0

This release provides the official implementation of MemSFT: Mitigating
Alignment Tax with an External Parametric Memory.

• Paper: https://arxiv.org/abs/2607.25614
• Hugging Face collection: https://huggingface.co/collections/Jiarui-Wang/memsft
• Checkpoint manifest: https://github.com/LUMIA-Group/MemSFT/blob/v1.0.0/docs/release/checkpoint_release_manifest.json
• SHA256 checksums: https://github.com/LUMIA-Group/MemSFT/blob/v1.0.0/docs/release/checkpoints.sha256

Highlights

• Releases the MemSFT training, routing, generation, and evaluation code under
  the memsft Python package.
• Provides Biology-Instructions, OpenSWI, and Law memory checkpoints, together
  with 11 compatible router checkpoints in tensor-only safetensors format.
• Includes the ordered OpenSWI shallow-1K evaluation split and a complete
  Qwen3-14B + Qwen3-8B OpenSWI reproduction path.
• Adds a tested OpenSWI generation example and an automated release audit.
• Aligns the source package, citation metadata, Git tag, and checkpoint release
  revision at 1.0.0 / v1.0.0.

The release assets include copies of the checkpoint manifest and SHA256 list.
Model weights remain hosted in the Hugging Face collection linked above.