Abdullahu5mani/flowscribe-qwen2.5-0.5b-v2

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:0.5BQuant:BF16Ctx Length:32kPublished:Apr 2, 2026License:mitArchitecture:Transformer Open Weights Warm

Abdullahu5mani/flowscribe-qwen2.5-0.5b-v2 is a 0.5 billion parameter Qwen2.5-Instruct model fine-tuned by Abdullahu5mani for speech-to-text post-processing. This model specializes in converting raw, messy speech transcripts into clean, formatted text across various writing styles, including Professional, Casual, and Software_Dev. With a 32768 token context length, it efficiently cleans filler words, corrects grammar, and applies structural formatting to improve transcript readability and utility.

Loading preview...

FlowScribe: Qwen2.5-0.5B Speech Transcript Formatter

Abdullahu5mani/flowscribe-qwen2.5-0.5b-v2 is a specialized fine-tuned version of the Qwen2.5-0.5B-Instruct model, designed to address the common issues with raw speech-to-text (STT) outputs. Tools like Whisper often produce transcripts filled with filler words, self-corrections, and lacking proper punctuation or formatting. This model acts as a post-processor, transforming these raw inputs into polished, readable text.

Key Capabilities

  • Intelligent Text Formatting: Removes filler words ("um", "uh"), handles self-corrections, fixes grammar, and applies appropriate punctuation and structure.
  • Multi-Style Output: Supports various formatting styles to match specific use cases:
    • Auto: An intelligent default for general cleanup and structuring.
    • Professional: Formal tone, structured layout, perfect grammar.
    • Casual: Preserves speaker's voice, light cleanup, contractions.
    • Verbatim: Strips only "um"/"uh" and applies spoken formatting commands.
    • Software_Dev: Formats code terms, variable names (e.g., camelCase, snake_case), and technical jargon.
    • Enthusiastic: High energy, exclamation marks, positive phrasing.
  • Efficient Processing: At 0.5 billion parameters, it prioritizes speed and local deployment, making it suitable for resource-constrained environments.
  • Quantized Version Available: Includes a Q4_K_M quantized GGUF version for fast CPU/GPU inference via llama.cpp.

Training Details

FlowScribe was fine-tuned using LoRA on approximately 27,400 synthetically generated examples from the flowscribe-dataset. This dataset was created using Google Gemini and other OpenRouter models across 10 diverse domain scenarios, ensuring a broad understanding of formatting requirements.

Limitations

  • Optimized exclusively for English language processing.
  • Relies on synthetically generated training data, which may not cover all real-world dictation edge cases.
  • The small parameter size (0.5B) prioritizes speed and local deployment over raw linguistic capability, meaning complex reasoning tasks are not its primary focus.

Finetunes
1 models