IntDim-L 50% pruned google/gemma-4-26B-A4B

This checkpoint was structurally pruned with the released Less-is-MoE mean-absolute-gradient method. It removes exactly 50% of routed-expert FFN neurons using 64 calibration samples from the gpqa_main configuration of Idavidrein/gpqa revision 633f5ee89ab8ad4522a9f850766b73f62147ffdd. The frozen settings are train, seed 1234, question plus shuffled choices, expert Explanation, answer letter, no chat template, no tokenizer length limit, no truncation, no padding, BF16 gradient scoring, and no optimizer step. The complete pretrained Gemma 4 causal language-model tower was loaded for FP32 equivalence validation and saved in BF16; the vision tower is intentionally excluded from this text-generation checkpoint.

The source-row selection hash is 790c4c22309def44542965fdde7c5f38f1d8e354602640cfb31518134b8d92e6 and the Gemma token-file hash is 1f160f6cebc16683000226e0456ce03113c5304830cb575b80ed3f4521abe2e0. The frozen source selection is recorded in jayzou3773/less-is-moe-gpqa-main-calibration-64 revision 7134dfef5af4605eae0706c30efa9226f49aed96. The Gemma token tensors were not added to that repository; their hash and per-sample token hashes are recorded without publishing another copy of GPQA-derived content. Full export and zero-mask equivalence metadata are in experiment-export.json.

Inference requires the Less-is-MoE ragged vLLM plugin from the unified Less-is-MoE GPU image. IntDim-E has one uniform expert width. IntDim-L/G retain the routed MoE topology and store compact per-expert widths in config.json.