Gemma 4 E4B IT Coder
This is the full merged coding-tuned model from google/gemma-4-E4B-it plus the
LoRA adapter josephmayo/gemma-4-E4B-it-coding-lora.
The release is not adapter-only: the LoRA deltas were merged directly into the
base safetensors and uploaded as a normal Transformers model.
Training Proof
Training ran on Kaggle with 2x Tesla T4 GPUs.
| Item | Value |
|---|---|
| Safe coding rows | 1024 |
| LoRA steps | 200 |
| LoRA rank | 16 |
| LoRA alpha | 32 |
| Trainable parameters | 50,499,584 |
| Final train loss | 1.1427 |
| Merged LoRA tensors applied | 592/592 |
| Missing LoRA targets | 0 |
| Merged safetensor shards | 5 |
HumanEval Results (50-Problem Subset)
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 included in this repo:
eval50_summary.json: 50-problem HumanEval executable result.eval50_before_after_full_code.csv: full generated before/after code for all 50 tasks.EVAL50_README.md: evaluation methodology and scope.nvidia_smi.txt: GPU environment proof.eval_before_after.csv: fixed before/after coding prompt scores with output previews.trainer_log_history.json: training loss and runtime logs.merge_manifest.json: direct merge record, including 592 applied LoRA tensors and 0 missing targets.model.safetensors.index.json: shard index for the full merged model.
This model is for benign coding assistance only. The training filter removed malware, phishing, exploit, credential theft, evasion, and destructive automation examples.
Evidence files
Run evidence for this release is stored in the repository under evidence/:
These files are compact local/Kaggle run artifacts used to document training, evaluation, merge, or quantization evidence for this model family.