ReFyneLabs/simplewords-dictation-cleanup-v3-adapter

simplewords-dictation-cleanup-v3 — LoRA adapter

The rank-16 LoRA adapter for abhiram3040/simplewords-dictation-cleanup-v3. Full metrics, evaluation method and deployment notes are on the fused model's card.

Use the fused model unless you specifically need the adapter. It is behaviorally identical (verified: failure sets match exactly) and faster — a runtime adapter adds per-layer overhead a fused model does not have.

This adapter is published in its own repo on purpose. mlx-swift-lm enumerates a model directory recursively and merges every .safetensors it finds, then rejects unused keys — an adapter/ subfolder inside the fused repo fails the load with unhandledKeys([... "lora_a", "lora_b"]).

python -m mlx_lm.generate \
  --model mlx-community/Qwen3.5-2B-MLX-4bit \
  --adapter-path . --prompt "$(cat system_v2.txt)

let's meet thursday no friday"

If you fuse it yourself, use --dequantize → bf16 → re-quantize to 8-bit. A plain 4-bit fuse rounds the low-rank delta away and collapses self-correction accuracy (measured on v2: 100% → 57.8%) while validation loss still looks healthy.