---
base_model: mlx-community/gemma-4-12B-it-qat-4bit
license: gemma
library_name: mlx-vlm
tags:
  - lora
  - qlora
  - gemma4
  - agentic
  - coding
  - tool-use
  - mlx
datasets:
  - greghavens/fable-5-coding-and-debugging-traces
  - greghavens/gpt-5.6-sol-coding-and-debugging-traces
  - greghavens/kimi-k3-coding-and-debugging-traces
---

# gemma-4-12b-it-qat-4bit-frontierdistill

A rank-8 LoRA adapter for [mlx-community/gemma-4-12B-it-qat-4bit](https://huggingface.co/mlx-community/gemma-4-12B-it-qat-4bit), fine-tuned for agentic coding and tool use. Trained on Apple Silicon (MLX) via [mlx-vlm](https://github.com/Blaizzy/mlx-vlm).

## Training data

16,730 examples combined from three frontier-model distillation trace datasets (all CC-BY-4.0):

- [greghavens/fable-5-coding-and-debugging-traces](https://huggingface.co/datasets/greghavens/fable-5-coding-and-debugging-traces)
- [greghavens/gpt-5.6-sol-coding-and-debugging-traces](https://huggingface.co/datasets/greghavens/gpt-5.6-sol-coding-and-debugging-traces)
- [greghavens/kimi-k3-coding-and-debugging-traces](https://huggingface.co/datasets/greghavens/kimi-k3-coding-and-debugging-traces)

Split by trajectory (not row) so no trajectory leaks across train/valid/test.

## Training config

| | |
|---|---|
| Method | LoRA, all `language_model` linear layers |
| Rank / alpha | 8 / 16 |
| Iterations | 1000 (~1.1 epochs) |
| Effective batch | 16 (grad accumulation) |
| Learning rate | 2e-5, constant |
| Max sequence length | 8192 |
| Loss | response-only (prompt masked) |

## Evaluation

Held-out test split (never seen in training), base vs. this adapter:

| Metric | Base | +LoRA | Δ |
|---|---|---|---|
| Perplexity (held-out NLL) | 4.23 | 2.15 | −49% |
| HumanEval pass@1 (n=164) | 92.1% | 89.0% | −3.1pp |
| Tool-name accuracy (n=150) | 68.0% | 74.0% | +6.0pp |
| Tool-argument overlap (n=150) | 23.7% | 41.3% | +17.7pp |

The adapter meaningfully improves tool selection and argument correctness — the actual training objective — at a small cost to isolated, out-of-distribution coding-puzzle accuracy (HumanEval). Perplexity alone overstates the improvement; treat it alongside the functional numbers, not in place of them.

## Usage

```bash
pip install mlx-vlm
python -m mlx_vlm.generate \
    --model mlx-community/gemma-4-12B-it-qat-4bit \
    --adapter-path gemma-4-12b-it-qat-4bit-frontierdistill \
    --prompt "Your prompt here"
```