Fine-tuning Gemma 4 edge models lifts MATLAB signal-code pass rates
Original title: Fine-Tuning Gemma 4 Edge Models for MATLAB Signal-Processing Code with SFT and QLoRA on a Laptop iGPU
On 66 held-out checks, E2B rose from 18 to 37 passes and E4B QLoRA from 32 to 47; adapters cut speed 17-19%.
Researchers applied supervised fine-tuning to Gemma 4 edge models E2B and E4B using 66 checked MATLAB signal-processing examples, with LoRA and QLoRA on a laptop iGPU. Int8 models with adapters were deployed through OpenVINO. Held-out pass rates rose for both sizes. A review of 159 failing programs found the adapters removed most convention errors but few errors in building signals and noise.
- The 66 examples cover 11 MATLAB signal-processing subtasks.
- E2B reached 37 of 66 passes with both LoRA and QLoRA.
- E4B with QLoRA reached 47 of 66 passes, up from 32.
- Applied at run time, the adapter costs 17-19% of generation speed.
- Products
- OpenVINO
- AI models
- Gemma 4 E2BGemma 4 E4B
From the original article
1 min read at arxiv.orgSmall language models that run on a laptop are attractive for writing signal-processing code offline, but they make systematic errors on digital signal processing (DSP) conventions.
By Mehmet Altan Toksöz · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.