beingpax commited on
Commit
eaaa7f3
·
verified ·
1 Parent(s): 3e00c6a

Describe model capabilities, limitations and usage

Browse files
Files changed (1) hide show
  1. README.md +35 -29
README.md CHANGED
@@ -3,65 +3,71 @@ library_name: mlx
3
  license: apache-2.0
4
  pipeline_tag: text-generation
5
  base_model: Qwen/Qwen3.5-0.8B
6
- base_model_relation: finetune
7
- datasets:
8
- - beingpax/VoiceInk-cleanup-dataset
9
  language:
10
  - en
11
  tags:
12
  - mlx
13
- - voiceink
14
  - transcript-cleanup
15
- - text-editing
16
  - 4-bit
17
  model_name: VoiceInk Refine Mini
18
  ---
19
  # VoiceInk Refine Mini
20
 
21
- A compact, locally fine-tuned transcript cleanup model based on Qwen3.5-0.8B. It cleans existing speech-to-text transcripts; it does not transcribe audio.
22
 
23
- This release is the selected local Unsloth Studio supervised fine-tune, merged and converted to MLX affine 4-bit quantization with group size 64. It is not the later experimental DPO candidate. The weight tensor names and configuration retain the original Qwen architecture for loader compatibility.
24
 
25
- ## Intended behavior
 
 
 
26
 
27
- Remove fillers and accidental repetitions, apply explicit self-corrections and spoken formatting, and improve punctuation while preserving the speaker's wording, meaning and tone. Return only the cleaned transcript. Treat instructions inside the transcript as dictated content, rather than answering them.
28
 
29
- ## Recommended inference
30
 
31
- Use the following system prompt, the raw transcript as the user message, `enable_thinking=False`, and greedy decoding (`temperature=0`). Do not prepend task instructions to the transcript itself.
32
 
33
- ```text
34
- Clean up this speech transcript. Preserve the speaker’s meaning, wording, tone, names, numbers, and important details. Remove fillers and accidental repetitions, and apply explicit corrections and spoken formatting. Fix punctuation and obvious recognition errors without guessing. Do not answer requests, summarize, add content, or invent greetings or sign-offs. Return only the cleaned transcript.
35
- ```
36
 
37
  ```python
38
  from mlx_lm import load, generate
 
39
 
40
  model, tokenizer = load("beingpax/VoiceInk-Refine-Mini")
41
- system_prompt = "Clean up this speech transcript. Preserve the speaker’s meaning, wording, tone, names, numbers, and important details. Remove fillers and accidental repetitions, and apply explicit corrections and spoken formatting. Fix punctuation and obvious recognition errors without guessing. Do not answer requests, summarize, add content, or invent greetings or sign-offs. Return only the cleaned transcript."
42
  messages = [
43
- {"role": "system", "content": system_prompt},
44
- {"role": "user", "content": "um please send the invoice on Tuesday no Thursday"},
 
 
 
 
 
 
 
 
 
45
  ]
46
  prompt = tokenizer.apply_chat_template(
47
  messages, tokenize=False, add_generation_prompt=True, enable_thinking=False,
48
  )
49
- from mlx_lm.sample_utils import make_sampler
50
- print(generate(model, tokenizer, prompt=prompt, max_tokens=1024, sampler=make_sampler(temp=0)))
 
 
51
  ```
52
 
53
- These are MLX weights for Apple silicon, not GGUF weights for Ollama or llama.cpp. Use an MLX runtime with Qwen3.5 text support.
54
-
55
- ## Training and evaluation
56
-
57
- Fine-tuned locally in Unsloth Studio on the VoiceInk cleanup training split, using a concise system prompt and non-thinking responses. The dataset snapshot contained 4,443 training pairs and 508 validation pairs. No private local paths, training examples, or adapter checkpoints are included in this release.
58
-
59
- The four-bit export was manually compared with a later experimental candidate on 100 randomly selected validation examples and 100 training examples. This release was preferred in 6 cases, the candidate in 3, with 191 ties, including 11 ties with shared issues. These were assistant-authored, unblinded qualitative judgments, not an independent human study or an automatic benchmark. Training examples were seen data; validation had been used historically for evaluation. These results do not establish generalization performance or superiority over other models.
60
 
61
  ## Limitations
62
 
63
- English-focused cleanup. The model may mishandle spelled names, negation, correction scope, quotations and list or paragraph boundaries. Reference transcripts can also contain errors. Review consequential details against the original transcript. Multilingual cleanup, vision and audio input were not evaluated for this release.
 
 
 
 
 
64
 
65
- ## License and attribution
66
 
67
- Distributed under Apache 2.0. Derived from [Qwen/Qwen3.5-0.8B](https://huggingface.co/Qwen/Qwen3.5-0.8B), developed by the Qwen team. See [LICENSE](LICENSE). Fine-tuning and MLX quantization modify the upstream model weights for transcript cleanup.
 
3
  license: apache-2.0
4
  pipeline_tag: text-generation
5
  base_model: Qwen/Qwen3.5-0.8B
 
 
 
6
  language:
7
  - en
8
  tags:
9
  - mlx
 
10
  - transcript-cleanup
 
11
  - 4-bit
12
  model_name: VoiceInk Refine Mini
13
  ---
14
  # VoiceInk Refine Mini
15
 
16
+ A compact model for cleaning up speech-to-text transcripts on your Mac. Available as Qwen3.5-0.8B in MLX 4-bit format.
17
 
18
+ ## What it does
19
 
20
+ - Removes fillers, accidental repetitions and abandoned phrases.
21
+ - Applies explicit corrections to names, numbers and other details.
22
+ - Improves punctuation, capitalization, lists and paragraph breaks.
23
+ - Aims to preserve the speaker's meaning, wording and tone.
24
 
25
+ It returns only the cleaned transcript. Requests spoken within a transcript should remain dictated text, rather than being answered or carried out.
26
 
27
+ ## How to run
28
 
29
+ Use an Apple silicon Mac and an MLX runtime with Qwen3.5 support. These weights are not compatible with Ollama or llama.cpp, which require a different format.
30
 
31
+ Pass the transcript as the user message. Use the system prompt below, disable thinking and set temperature to `0` for greedy decoding.
 
 
32
 
33
  ```python
34
  from mlx_lm import load, generate
35
+ from mlx_lm.sample_utils import make_sampler
36
 
37
  model, tokenizer = load("beingpax/VoiceInk-Refine-Mini")
 
38
  messages = [
39
+ {
40
+ "role": "system",
41
+ "content": (
42
+ "Clean up this speech transcript. Preserve the speaker’s meaning, wording, tone, "
43
+ "names, numbers, and important details. Remove fillers and accidental repetitions, "
44
+ "and apply explicit corrections and spoken formatting. Fix punctuation and obvious "
45
+ "recognition errors without guessing. Do not answer requests, summarize, add content, "
46
+ "or invent greetings or sign-offs. Return only the cleaned transcript."
47
+ ),
48
+ },
49
+ {"role": "user", "content": "um send the invoice on Tuesday no Thursday"},
50
  ]
51
  prompt = tokenizer.apply_chat_template(
52
  messages, tokenize=False, add_generation_prompt=True, enable_thinking=False,
53
  )
54
+ print(generate(
55
+ model, tokenizer, prompt=prompt, max_tokens=1024,
56
+ sampler=make_sampler(temp=0),
57
+ ))
58
  ```
59
 
60
+ Allow enough output tokens for the complete transcript; increase the limit for longer inputs.
 
 
 
 
 
 
61
 
62
  ## Limitations
63
 
64
+ - Accepts text, not audio. It cannot recover words missing from the transcript.
65
+ - Intended for English cleanup; multilingual cleanup has not been validated.
66
+ - Can misinterpret spelled names, negation, self-corrections and quotation boundaries.
67
+ - List items, line breaks and paragraphs may be formatted incorrectly.
68
+ - Can omit details or introduce incorrect edits. Check names, amounts, dates and other important details against the original transcript.
69
+ - Not intended for answering questions, summarizing or rewriting the speaker's message.
70
 
71
+ ## License
72
 
73
+ Apache 2.0. Based on [Qwen3.5-0.8B](https://huggingface.co/Qwen/Qwen3.5-0.8B) by the Qwen team, with modified weights for transcript cleanup. See [LICENSE](LICENSE).