amd/gemma-3-4b

• Introduction

  This model was created using Quark Quantization for the Decoder, followed by OGA Model Builder, and finalized with post-processing for NPU deployment.

• Quantization Strategy

  • AWQ / Group 128 / Asymmetric / BFP16 activations / UINT4 weights

• Base model info:

• Please refer to Gemma-3-4b-it for base model info.

Evaluation scores

• The MMMU scores are, Music: 33.33, Marketing: 40, Math: 30, Clinical Medicine: 20, History: 40, Electronics: 13.3.
• The perplexity measurement is run on the wikitext-2-raw-v1 (raw data) dataset provided by Hugging Face. Perplexity score measured for prompt length 2k is 16.825.

License