deutsch_a2_transformer

Deutsch A2 Grammar Transformer (Ukrainian Explanations)

This is a compact, custom-built BART (Encoder-Decoder) Transformer model designed to identify, correct, and explain German grammar errors at the A1-A2 level. It is specifically tailored for Ukrainian-speaking learners, providing detailed grammar feedback and explanations in Ukrainian.

✨ Live Demo (Spaces): kengurukleo/deutsch-a2-tutor

🚀 Model Capabilities

The model covers over 18 essential grammar topics for A1 and A2 levels.

📚 Covered Topics & Examples

🟢 Level A1: Basic Grammar

🟡 Level A1/A2: Intermediate Topics

🔵 Level A2: Advanced Grammar

🛠 Architecture

• Type: Transformer Encoder-Decoder (BART-style)
• Parameters: ~12M
• Encoder Layers: 3
• Decoder Layers: 3
• Attention Heads: 4
• Embedding Dim: 256
• Tokenizer: Byte-level BPE (8,000 source tokens, 8,003 total)

📖 How to Use

The model is fully compatible with the standard transformers library. No custom code is required.

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer, pipeline

# Load model and tokenizer
model_id = "kengurukleo/deutsch_a2_transformer"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSeq2SeqLM.from_pretrained(model_id)

# Use pipeline for easy inference
corrector = pipeline("text2text-generation", model=model, tokenizer=tokenizer)
result = corrector("Ich habe nach Berlin gefahren.")
print(result[0]['generated_text'])

📜 Release History

• v1.0 (Proof of Concept): Initial release with a basic Transformer Decoder and word-level vocabulary.
• v1.1 (BPE Tokenizer): Migration to a Byte-level BPE tokenizer (8,000 tokens).
• v1.2 (GPT-2 Native): Refactored to be a fully compatible GPT-2 model.
• v2.0 (Encoder-Decoder): Major upgrade to a BART-style architecture. Switched to Seq2Seq modeling for superior grammatical logic and explanations.

👨‍💻 Author

Created by KenguruKleo.