Jeromeof/pragmatic-audio-gemma4-e4b-mlx-adapter

Pragmatic Audio Gemma 4 E4B — experimental MLX adapter

Not suitable for deployment — training defect identified on 7 September 2026. The training path supplied a padding mask that bypassed Gemma 4's causal attention-mask construction. A prefix-invariance check confirmed future-token leakage. The final adapter also repeated a product name on a held-out question; the 100-step checkpoint repeated words on another held-out question. The original training loss is not evidence of valid causal language learning. These weights are retained as an experimental artifact, not a recommended model. A corrected training run and fresh evaluation are required.

An experimental LoRA adapter trained for source-grounded audio discussion. This repository contains adapter weights, not a standalone model or an ONNX browser model. Load it with the matching base model using MLX-VLM.

Contents

• adapters.safetensors: final 1,000-step adapter.
• adapter_config.json: MLX LoRA configuration.
• training_summary.json: training settings and limitations.

Training

Trained on 94 conversational examples from an audio-discussion dataset with source provenance, including GPT-assisted annotations. Six examples were held out. Training used batch size 1, a maximum sequence length of 2,048, learning rate 0.00002, and gradient checkpointing. Training completed on 6 September 2026. Raw training data is not included. Source-ID auditing found 90 review examples, 3 Audio Science Review forum examples, and 1 Headphones.com forum example. None of the separately extracted book/paper references entered this training set. Improved audio-science alignment has not been demonstrated.

Evaluation and limitations

Held-out answer-quality evaluation has not been performed. Successful training and low training loss do not establish improved accuracy. The final checkpoint may overfit this small dataset. Audio claims require verification against their sources. The original run did not pin a base revision. The sole local cached snapshot and the cached main reference inspected after training both identify 475b9088d29754a3379866cf5aeb6b41acd313c2. This is the inferred training base revision, not a revision recorded at training time.

Usage

Use MLX-VLM on a compatible Apple Silicon machine. Pass this downloaded repository directory as adapter_path when loading mlx-community/gemma-4-e4b-it-4bit. Inference with this final adapter has not yet been validated. Browser use requires a separate merged, converted ONNX release.

Release status and upstream licensing

Experimental MLX adapter release; not browser-ready. No held-out quality claims are made. This repository does not include the base weights or training corpus.

The MLX base model card labels its license gemma, while the original Google model card labels its license apache-2.0. These upstream metadata differ. No adapter license is assigned by this release; do not interpret publication as a new license grant.

Download

hf download Jeromeof/pragmatic-audio-gemma4-e4b-mlx-adapter --local-dir ./pragmatic-audio-adapter

Load the base model and pass ./pragmatic-audio-adapter as its MLX-VLM adapter path. This is an MLX adapter, not a PEFT adapter or an ONNX model.