Buttermilk03/scriber-gemma3-270m-polishing-de-v1

TEXT GENERATIONConcurrent Unit Cost:1Model Size:0.3BQuant:BF16Context Size:32kPublished:Aug 1, 2026License:gemmaArchitecture:Transformer Featherless Exclusive Cold

Buttermilk03/scriber-gemma3-270m-polishing-de-v1 is a 0.3 billion parameter model, fine-tuned from google/gemma-3-270m-it, specifically designed for conservative local transcript cleanup into Scriber Structured Text v1. This model specializes in polishing speech-to-text transcripts, focusing on maintaining protected spans and requiring fail-closed validation. It is intended for post-processing existing transcripts rather than generating them from audio.

Loading preview...

Model Overview

Buttermilk03/scriber-gemma3-270m-polishing-de-v1 is an experimental, private candidate model developed by Buttermilk03. It is a full fine-tune of google/gemma-3-270m-it, specifically tailored for the task of conservative local transcript cleanup into Scriber Structured Text v1. This model is not a speech-to-text (STT) model itself but rather a post-processing tool for existing transcripts.

Key Characteristics

  • Transcript Polishing: Optimized for refining and cleaning up raw transcripts.
  • Scriber Structured Text v1: Designed to output cleaned transcripts in a specific structured text format.
  • Protected-Span Handling: Requires careful handling of protected spans within the text to prevent unintended alterations.
  • Fail-Closed Validation: Emphasizes robust validation, recommending rejection of invalid output, unknown tags, or altered protected spans, and falling back to original text if validation fails.
  • Experimental Status: Currently a PRIVATE_CANDIDATE, indicating it is experimental and not a production release, with quality gates having failed in automatic evaluations.

Usage Considerations

This model should be used with deterministic greedy generation and requires adherence to a bundled runtime contract. Users must implement validation to reject outputs that contain invalid Scriber Structured Text, unknown tags, changed protected spans, or invented structures. It is crucial to fall back to conservative plain text or the original transcript if validation issues arise. The model is subject to Gemma's Terms of Use and Prohibited Use Policy and does not provide specialized advice (e.g., legal, medical).