josephmayo/gemma-4-E4B-it-coding-lora

Gemma 4 E4B IT Coding LoRA

QLoRA adapter for google/gemma-4-E4B-it, trained on filtered benign coding instructions.

Training

• Runtime: Kaggle 2x Tesla T4
• Data: filtered benign coding instruction data
• Safe rows used: 1024
• Steps: 200
• LoRA: r=16, alpha=32, target_modules=all-linear
• Trainable parameters: 50,499,584
• Final train loss: 1.1427

50-Problem HumanEval Proof

This adapter was merged into josephmayo/gemma-4-E4B-it-Coder and 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 merge
Pass count	34 / 50	42 / 50
Absolute lift	-	+16.0 pp
Relative pass-count lift	-	+23.53%

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

Earlier 8-task smoke artifacts are still included for reproducibility (eval_before_after.csv, executable_eval.json, trainer_log_history.json, summary.json, proof_summary.json, evaluation_scope.json), but the headline proof is the 50-task executable run above.

This adapter is for benign coding assistance only. It was not trained on malware, phishing, exploit, credential theft, evasion, or destructive automation examples.