| --- |
| 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. |
|
|
