spanthee/gemma4-e4b-macos-clawdia-toolcalling-mlx-lora

Gemma 4 E4B macOS Clawdia Tool-Calling MLX LoRA

MLX LoRA adapter for Gemma 4 E4B, trained on the same macOS, Clawdia, Cua Driver, schema-conditioned, and local-agent tool-calling data used for the Qwen tool-calling run.

This repository is an adapter-only artifact. Load it with a compatible Gemma 4 E4B MLX base model rather than treating it as a fused model.

Local MLX usage

python -m mlx_lm chat \
  --model mlx-community/gemma-4-e4b-it-4bit \
  --adapter-path spanthee/gemma4-e4b-macos-clawdia-toolcalling-mlx-lora \
  --temp 0.1 \
  --top-p 0.9 \
  --max-tokens 1200

For tool-calling use, include the available tool names, schemas, and tool policy in the prompt. The adapter is trained to bind user intent to the current tool list rather than relying on fixed tool names.

Eval snapshot

Text/structured tool-calling evals are currently strong in the local harness: focused structured and unavailable-tool suites pass. Multimodal receipt-image diagnostics remain weak for this artifact, so do not treat this adapter as ready for direct image receipt ingestion without another VLM-specific pass.

Adapter details

• MLX adapter rank: 64
• MLX adapter alpha: 128
• Converted from PEFT LoRA trained from google/gemma-4-E4B-it