murmur-cleanup-0.5b

Dictation cleanup model for Murmur, the private hold-to-talk dictation app. It turns raw speech-to-text into the text you meant: it applies self-corrections ("Tuesday, no wait, Thursday"), removes fillers, fixes punctuation, writes numbers as digits, and never answers or adds anything.

Fine-tuned from Qwen/Qwen2.5-0.5B-Instruct with LoRA on 19,931 generated dictations.

Murmur benchmark (held-out test dictations): meaning correct 99.3%, format 100.0%.

Use in Murmur: set "llamacpp_model": "YungKeegs/murmur-cleanup-0.5b:murmur-cleanup-0.5b-q8_0.gguf" in settings.

Downloads last month
63
GGUF
Model size
0.5B params
Architecture
qwen2
Hardware compatibility
Log In to add your hardware

8-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for YungKeegs/murmur-cleanup-0.5b

Quantized
(397)
this model