stukenov/sozkz-fix-qwen-500m-kk-gec-v4

TEXT GENERATIONConcurrency Cost:1Model Size:0.5BQuant:BF16Ctx Length:32kPublished:Apr 20, 2026License:mitArchitecture:Transformer Open Weights Gated Cold

The stukenov/sozkz-fix-qwen-500m-kk-gec-v4 is a 447 million parameter Kazakh grammatical error correction model, fine-tuned by stukenov using Kahneman-Tversky Optimization (KTO). This model specifically enhances punctuation handling, including comma and period insertion, and corrects grammar and word usage errors in Kazakh text. It is designed to be integrated into a larger pipeline with an 'emle' (spelling) pre-fixer for comprehensive error correction.

Loading preview...

SozKZ Fix Qwen 500M — Kazakh GEC v4 Overview

This model, developed by stukenov, is a 447 million parameter Kazakh grammatical error correction (GEC) model. It is the fourth version in its series, building upon stukenov/sozkz-fix-qwen-500m-kk-gec-v3.

Key Differentiators & Capabilities

  • KTO Preference Optimization: Utilizes Kahneman-Tversky Optimization (KTO) on 26,404 preference pairs to learn output preferences, specifically improving punctuation handling beyond standard supervised fine-tuning.
  • Enhanced Punctuation Correction: Significantly improves comma insertion before conjunctions and after introductory words, as well as period placement, through dedicated training data.
  • Grammar and Word Usage Correction: Addresses general grammatical and word usage errors in Kazakh text.
  • Pipeline Integration: Designed for optimal performance when used with an external 'emle' (spelling) pre-fixer, which handles character substitution errors, allowing the GEC model to focus on grammar and punctuation.

Use Cases

  • Automated Kazakh Text Correction: Ideal for applications requiring automated correction of grammatical, punctuation, and word usage errors in Kazakh language content.
  • Improving Text Quality: Useful for enhancing the readability and correctness of written Kazakh, particularly in contexts where precise punctuation is critical.

Limitations

  • Standalone 'emle' (spelling) accuracy is reduced; it requires the external 'emle' pipeline for comprehensive spelling correction.
  • The model's standalone performance on a custom GEC benchmark is 5%, indicating its specialized role within a broader correction pipeline rather than as a standalone general-purpose fixer.