ParrotFlow term judge
A LoRA adapter for ParrotFlow, a local dictation app for macOS. It answers one question about a dictated sentence: should a misheard word become a term from the user's own list?
This revision answers the v2 prompt. One Term: line per candidate, with an optional definition after —, then the sentence with the place in brackets:
Term: Zod — TypeScript schema validation library
Sentence: we validate the payload with [sod] before saving
Change "sod" to "Zod"?
The score is P(" yes") against P(" no") at the last position. Terms and definitions are given at run time: the adapter learned the skill on made-up names, not on any real list.
The earlier revision ef04c0379b37db81d883dcfc7ad3b744fe64b907 answers the v1 prompt (Terms: list, no brackets). App versions that pin it keep working.
Base model
The adapter only works with this exact base, the 4-bit MLX conversion of Qwen/Qwen3-1.7B-Base:
- repo
bluevisor/Qwen3-1.7B-Base-4bit-mlx, revisiond12621072e81eb70d6ba1f01583141785c74fa6a model.safetensorssha2563b29904336e6b2fdfb80af60541e23a14b3e729de80df0ca178f661a9d0f448d
The same file comes out of mlx_lm convert --hf-path Qwen/Qwen3-1.7B-Base -q --q-bits 4 (mlx-lm 0.32.2, mlx 0.32.4) from Qwen/Qwen3-1.7B-Base revision ea980cb0a6c2ae4b936e82123acc929f1cec04c1.
Training
- LoRA rank 8, scale 20, last 16 layers, learning rate 7e-5, batch 4, seed 1, epoch 2 of 3.
- 3,688 training rows and 390 validation rows, all invented: texts written by language models around made-up names, ordinary words used as product names, acronyms written as words, and people's names. About 16% French. Sentences were voiced with local text-to-speech and transcribed, so the misheard spans are real recognizer output. No real user dictation.
- Validation accuracy 0.936 at epoch 2.