mumble-cleanup-dataset

A hand-curated synthetic dataset of (raw, clean) dictation pairs used to fine-tune adikuma/mumble-cleanup, a small LM that cleans ASR output into properly-formatted prose.

raw mimics how an ASR system (Parakeet) actually produces text: lowercase, no punctuation, no apostrophes, fillers and stutters preserved. clean is what the speaker actually meant: proper capitalization, punctuation, contractions, no disfluencies, faithful to content.

At a glance

• Pairs: 688
• Language: English
• Categories: 8
• Generation: 8 specialist sub-agents wrote per-category batches; a polish pass added opener diversity and longer-form examples
• Mean faithfulness (Levenshtein on content words): 0.929

Schema

field	type	description
category	string	one of the 8 dictation categories below
raw	string	asr-style input: lowercase, no punctuation, fillers preserved
clean	string	the cleaned target: capitalized, punctuated, faithful

Category distribution

category	count
long_form_thoughts	145
casual_messages	81
meeting_notes	81
todo_lists	81
professional_emails	80
technical_dictation	80
questions_and_asks	70
mixed_content	70

Quick start

from datasets import load_dataset

ds = load_dataset("adikuma/mumble-cleanup-dataset", split="train")
print(ds[0])
# {"category": "casual_messages",
#  "raw": "hey sarah you free for lunch around noon",
#  "clean": "Hey Sarah, you free for lunch around noon?"}

Intended use

This dataset is designed for supervised fine-tuning of small instruction-tuned language models on the narrow task of dictation cleanup. The training recipe used for adikuma/mumble-cleanup:

• Base: Qwen/Qwen2.5-0.5B-Instruct
• Method: LoRA SFT, completion-only masking
• Split: stratified 85/10/5 train/val/test by category

The dataset is small enough to fine-tune on a single consumer GPU in minutes.

Limitations

• English only.
• Synthetic. Real ASR output may have failure modes the synthetic pairs do not capture. Sample distribution was tuned to match Parakeet's empirically observed failure pattern (no punctuation, lowercase, lost contractions, common disfluencies), but domain shift in production is expected.
• Faithfulness floor: legitimate transformations (e.g. "two thirty" → "2:30", "dont" → "don't", "us" → "U.S.") drive faithfulness below 1.0 even when the cleaning is correct. The mean of 0.929 reflects this, not noise.

License

Apache-2.0.

Citation

@misc{mumble-cleanup-dataset,
  title  = {Mumble Cleanup Pairs},
  author = {Kumar, Aditya},
  year   = {2026},
  url    = {https://huggingface.co/datasets/adikuma/mumble-cleanup-dataset}
}