itzune/gemma-4-e4b-horkonpon

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7.9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 31, 2026License:gemmaArchitecture:Transformer Featherless Exclusive Cold

The itzune/gemma-4-e4b-horkonpon model is a fine-tuned Google Gemma 4 E4B (8B total / 4.5B effective parameters) specifically for Basque (euskara) grammatical error correction. It generates structured JSON output for each correction, including category, normative reference, and a Basque-language explanation. This model excels at providing explainable GEC, allowing runtime filtering of errors by type, severity, or normative status, and significantly outperforms its base model and GECToR-v2 on F0.5 for Basque GEC.

Loading preview...

Model Overview

itzune/gemma-4-e4b-horkonpon is a specialized Google Gemma 4 E4B model, fine-tuned for Basque (euskara) Grammatical Error Correction (GEC). It leverages a 7.9 billion parameter architecture with a 32768 token context length, optimized for generating structured, explainable JSON outputs for corrections.

Key Capabilities & Differentiators

  • Explainable GEC: Provides detailed metadata for each correction, including category, nature (error/suggestion), ref (normative reference), and a Basque error_description.
  • Structured Output: Guarantees 100% JSON parse rate, enabling programmatic filtering of corrections based on type or severity at inference time.
  • Performance: Achieves an F0.5 score of 80.8 on the horkonpon-corpus evaluation set, outperforming GECToR-v2 (78.8 F0.5) and demonstrating significant gains in semantic categories like spelling, word-choice (zalantza), punctuation, and proper nouns.
  • Fine-tuning Impact: Ablation studies show fine-tuning dramatically improves precision (+84.5%), recall (+47.9%), and reduces false positives on clean text (from 97.2% to 8.6%) compared to the base model, teaching it "minimal-edit discipline."

Ideal Use Cases

This model is best suited for applications requiring robust and explainable Basque grammatical error correction, particularly where:

  • Runtime filtering of error types is beneficial (e.g., ignoring editorial suggestions or specific error categories).
  • Detailed explanations for corrections are needed for user feedback or educational purposes.
  • High accuracy in semantic error detection is prioritized over morphological errors (where GECToR-v2 still holds an edge).
  • Integration into writing assistants or language tools for Basque speakers.

Finetunes
1 models