tweeze lenses

Jacobian lenses compressed for tweeze, a model-behaviour workbench running on MLX.

solarkyle-31b-q8c

solarkyle's Gemma 4 31B-IT lens (solarkyle/jspace-lenses, gemma-4-31b-it/lens.pt at f5767d1), compressed to the layers tweeze reads, L27–L58, at 8 bits (MLX affine quantization, group size 64). Each layer's input scale is multiplied by a power of two so that the fp16 group scales stay out of the subnormal range. It targets block 59 and was fit on 100 WikiText-103 prompts.

solarkyle-31b-q8c/solarkyle-31b-q8c.safetensors: for each layer, L{l}.q (uint8 codes), L{l}.scales and L{l}.biases (fp16), and L{l}.s (fp32 input scale). 983,344,936 bytes, sha256 4a280218c660df60226a537dff1ee870b421e99a9b82f61b4ea7e01a43b9593e.
solarkyle-31b-q8c/manifest.json: the file's sha256, the source lens's metadata, and validation figures against the fp16 lens.

At the scored layers, L28–L56 and L58, its readouts differ from the fp16 lens's in 1.84% of cells, with 1 decisive miss in 75,946. L57 is unscored.

Licensed Apache 2.0, as the source lens is.

Downloads last month: -; Downloads are not tracked for this model. How to track

MLX

Hardware compatibility

Quantized

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for andyoneal/tweeze-lenses

Base model

google/gemma-4-31B

Finetuned

google/gemma-4-31B-it

Finetuned

(347)

this model