Evaluated on Kaggle with 2x Tesla T4 GPUs using an executable 50-task HumanEval subset. Full generated before/after code is published in eval50_before_after_full_code.csv.

Metric	Base `google/gemma-4-E4B-it`	Coder
Pass count	34 / 50	42 / 50
Absolute lift	-	+16.0 pp
Relative pass-count lift	-	+23.53%

Proof files: eval50_summary.json, eval50_before_after_full_code.csv, EVAL50_README.md, nvidia_smi.txt.

Local GGUF sanity check was run with llama.cpp b9133 on CPU. The Q3_K_M file loaded and generated a simple Python add(a, b) implementation. CPU speed was slow on this Windows machine, around 0.8 tokens/s, so use GPU llama.cpp, LM Studio, Ollama, or another accelerated runtime for normal use.

Example

llama-cli -m Gemma-4-E4B-it-Coder-Q5_K_M.gguf   -p "Write a Python function is_prime(n). Return only code."   -n 256 --temp 0.2 --ctx-size 2048

Notes

This repository publishes the quantized runtime artifacts only.
The merged HF safetensors model is available at josephmayo/gemma-4-E4B-it-Coder.
Training data details are intentionally summarized as filtered benign coding instruction data.

Evidence files

Run evidence for this release is stored in the repository under evidence/:

evidence/holykeys_gemma4_sft_output_release_release_summary.json

These files are compact local/Kaggle run artifacts used to document training, evaluation, merge, or quantization evidence for this model family.

josephmayo/gemma-4-E4B-it-Coder-GGUF

Gemma-4-E4B-it-Coder GGUF

Files

50-Problem HumanEval Proof

Example

Notes

Evidence files

josephmayo/
gemma-4-E4B-it-Coder-GGUF