gemma-4-31B-sparsegpt-unstructured-0.5

One-shot sparsegpt-unstructured pruned (ratio 0.5, actual 0.5) version of google/gemma-4-31B, produced as part of a minimal reproduction of the granularity-ordering mechanism in arXiv 2606.14150 (Small LLMs: Pruning vs Training from Scratch).

metric value
dense wikitext-2 ppl 5.6335
pruned wikitext-2 ppl 7.1283
Δ ppl 1.4947
calibration 64 samples @ seqlen 4096

Note: unstructured / N:M sparsity zeroes weights in place — the parameter count and file size are unchanged; this is an initialization-quality probe, not a size-reduction. See the paper for the granularity/hardware trade-off.

Downloads last month
11
Safetensors
Model size
31B params
Tensor type
BF16
·
BOOL
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for reneeice/gemma-4-31B-sparsegpt-unstructured-0.5

Finetuned
(56)
this model