localflow-clean

The cleanup model of localflow: it turns what you said into the text you meant. It removes fillers and false starts, applies spoken corrections ("at 5, no wait, at 6"), makes lists and headings from dictated structure, uses your vocabulary's spellings, and turns spoken file names into @file references.

localflow downloads it on first use and runs it inside its own process with llama.cpp. You don't need to install anything else.

Why it's small and fast

• English-only vocabulary. Qwen3.5 ships 248k tokens for many languages, and on a 0.8B model each token is a row of weights in the embedding and output layers. Pruned to the 32k that English needs (English text tokenizes exactly as before), the file went from 542 to 350 MB and the memory it takes loaded from ~810 to ~390 MB, with identical outputs.
• The rules are in the weights. Fine-tuned on ~4,600 dictation/cleanup pairs, it needs a one-line instruction instead of the ~3,900 tokens of rules and examples a general model is given, so there is far less to read before answering.
• 0.8B parameters, Q4_K_M, thinking off.

Results

On an RX 9070 XT (Vulkan), against qwen3.5:4b in Ollama, the model localflow used before, with localflow's full rules prompt, examples and a review pass.

	localflow-clean	qwen3.5:4b + review
localflow's written test cases (cases.txt)	32/41	35/41
Cleanup, under 30 words	56 ms	327 ms
Cleanup, 100+ words	338 ms	1,777 ms
Load from disk	~0.1 s	~2.5 s
GPU memory while loaded	~530 MB	~4.3 GB

Close to the 4B on the written cases, at a sixth of the time and an eighth of the memory. Where it still loses: corrections deep in a long ramble and repeated phrases ("this is this is this is").

On 142 held-out examples it matched the expected text exactly in 51% of cases (the 4B: 27%), but those come from the same distribution as its training data and picked the checkpoint, so they favor it.

Details

• Base: Qwen3.5-0.8B, LoRA fine-tuned (r 64, all linear layers).
• File: localflow-clean-q4_k_m.gguf, 350 MB. SHA-256 0fa7ba4e4a38b50067a2dbd6eb01ce2e0aa6db755afc4a216cffd27566eb3206.

Prompt format

The model expects the format it was trained on (Qwen chat template, thinking off). Omit the vocabulary block and <files> when you have none:

<|im_start|>system
Clean up the dictation.

Speaker's vocabulary (use these exact spellings):
- Pine Script: heard as "Py script"<|im_end|>
<|im_start|>user
<dictation>um so the py script for tradingview is uh done</dictation>
<files>
src/main.rs
</files><|im_end|>
<|im_start|>assistant
<think>

</think>

License

Apache 2.0, as for the Qwen3.5 base model (see LICENSE).