FrogBoss-35B-Flash / README.md
SofiTesfay2010's picture
Document custom all-STQ checkpoint
78c5697 verified
|
Raw
History Blame
1.12 kB
metadata
library_name: pytorch
base_model: microsoft/FrogBoss-32B-2510
tags:
  - quantized
  - custom-code
  - safetensors

All-STQ compressed checkpoint

Source: microsoft/FrogBoss-32B-2510 Source revision: cc930952b8751de86a2a44debaf26e45e677a291

This is a custom Sherry-style packed checkpoint, not a verified Tencent serialization format and not directly loadable with ordinary Transformers from_pretrained().

All supported weight matrices use 42 bytes per 256 weights: three signed active values per four coordinates, with one FP16 scale per 256 weights. Vectors and scalars remain FP16.

Importance weights are all one. No activation calibration, recovery training, or model-quality evaluation was performed.

Packed shard size: 5.377 GB.

The original architecture and parameter count are unchanged. Substantial accuracy loss and incoherent generations are possible.

Use the accompanying Colab custom-loader cells to run this representation. Retain a copy of that notebook: the packed tensors are not standard dense Transformers weights.

Review the original model's license before use or redistribution.