amitashwini/mumble-cleanup-2stage

TEXT GENERATIONConcurrent Unit Cost:1Model Size:0.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jun 23, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The amitashwini/mumble-cleanup-2stage model, developed by Echo Flow AI, is a 0.5 billion parameter language model fine-tuned from Qwen/Qwen2.5-0.5B-Instruct. It specializes in cleaning speech-to-text dictation transcripts by correcting capitalization, punctuation, and removing disfluencies without rephrasing or hallucinating content. This model was trained in a two-stage process, combining synthetic data with a small, hand-curated real-style dataset to achieve a 10/10 pass rate on the Echo Flow DictationQuality golden corpus.

Loading preview...

Model Overview

The amitashwini/mumble-cleanup-2stage model, developed by Echo Flow AI, is a 0.5 billion parameter language model based on Qwen/Qwen2.5-0.5B-Instruct. Its primary function is to clean raw speech-to-text dictation transcripts, transforming them into properly capitalized and punctuated text free of filler words and disfluencies.

Key Capabilities

  • Transcript Cleaning: Converts raw ASR output (lowercase, no punctuation, fillers) into clean, readable text.
  • Disfluency Removal: Eliminates filler words (e.g., "um", "uh"), stutters, repeated words, and false starts.
  • Punctuation and Capitalization: Automatically corrects punctuation and capitalization.
  • Non-Generative: Strictly adheres to a "no-reword, no-hallucination" contract, meaning it does not paraphrase, summarize, or add content not present in the original speech.
  • High Accuracy: Achieved a 10/10 pass rate on the Echo Flow DictationQuality golden corpus, outperforming its predecessor.

Training and Architecture

This model was fine-tuned using LoRA in a two-stage process:

  1. Stage 1 (Pretrain): Trained on 50,000 synthetic (raw, clean) pairs generated by the Echo Flow combinatorial template generator, covering diverse noise profiles and domains.
  2. Stage 2 (Fine-tune): Further refined on 638 hand-curated, real-style pairs from adikuma/mumble-cleanup-dataset with a lower learning rate to preserve its non-generative properties.

Use Cases

  • Post-processing ASR Output: Ideal for cleaning transcripts from speech-to-text systems for improved readability and accuracy.
  • Dictation Applications: Suitable for applications requiring clean, verbatim text from spoken input.

Limitations

  • English Only: Currently supports only the English language.
  • Input Length: Designed for short-to-medium dictation (up to approximately 512 tokens); longer inputs may require chunking.
  • Potential Over-correction: May occasionally over-correct when a user genuinely intends a fragmented sentence.

Finetunes
1 models