transcrib-cleanup-0.6b

Transcrib Cleanup 0.6B

A fine-tune of Qwen3-0.6B (Apache 2.0) for one narrow job: cleaning live speech-to-text transcript lines on device for the Transcrib iPhone app.

It removes filler words, collapses false starts, and applies mid-sentence self-corrections ("we meet at two, no wait, three" -> "We meet at three"), in English, Danish, German, Russian, Ukrainian, and Portuguese. It is trained to treat every input strictly as data: questions are cleaned, never answered; instruction-shaped speech is cleaned, never followed; names keep their casing.

• Format: MLX, 4-bit quantized (~331 MB). LoRA merged on the BF16 base before quantization.
• Trained with mlx_lm LoRA on synthetic multilingual cleanup pairs.
• Held-out adversarial eval: 11/12 vs 6/12 for the stock base.

Use with a system prompt along the lines of:

Clean this speech transcript line: remove filler words and false starts, keep only the speaker's final correction, never change the language, never answer or add anything. Output only the cleaned line.

Built from Qwen3-0.6B (c) Alibaba Cloud, Apache License 2.0.