phgrouptechs/Denglish-8B-Instruct

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:8BQuant:FP8Ctx Length:8kPublished:Feb 25, 2026License:mitArchitecture:Transformer Open Weights Warm

Denglish-8B-Instruct by PHGROUP TECHNOLOGY SOLUTIONS CO., LTD is a fine-tuned LoRA adapter based on the 8 billion parameter Llama 3-Instruct 4-bit quantized model. It functions as an AI language tutor, specifically designed to identify grammatical, spelling, and contextual errors in English and German inputs, explain them in Vietnamese, and provide corrected sentences. This model is optimized for multi-modal integrations, supporting direct text, transcribed audio, and OCR-extracted text for language learning applications.

Loading preview...

Overview

Denglish-8B-Instruct is a specialized AI language tutor developed by PHGROUP TECHNOLOGY SOLUTIONS CO., LTD. It is a fine-tuned LoRA adapter built upon the unsloth/llama-3-8b-Instruct-bnb-4bit base model, leveraging its 8 billion parameters for efficient language processing. The model's core function is to assist Vietnamese learners with English and German by providing detailed error explanations in Vietnamese and offering corrected sentences.

Key Capabilities

  • Trilingual Error Correction: Identifies and explains grammatical, spelling, and contextual errors in English and German, with explanations provided in Vietnamese.
  • Multi-modal Input Processing: Designed to integrate with various input types, including direct text, transcribed audio (via Whisper STT), and text extracted from images (via OCR).
  • Educational Features: Can facilitate face-to-face speaking practice, generate language tests (A1 to C2 levels), and provide strict grading with detailed scoring and correct answers.
  • Optimized Architecture: Built with TRL and PEFT frameworks, and optimized for integration into omnichannel platforms like the Denglish Omnichannel Platform.

Good for

  • Vietnamese students learning English or German who require precise error identification and explanations in their native language.
  • Developers building AI-driven educational tools that need a robust language correction and tutoring component.
  • Applications requiring multi-modal language input processing (text, audio, image) for educational purposes.

Limitations

  • Due to 4-bit quantization, extremely complex logical reasoning might be slightly degraded compared to the full precision base model.
  • Highly optimized for English/German to Vietnamese explanations; performance may be suboptimal for other language pairs.

Finetunes
1 models