🧠 Model Card: elimsoul-v1
🏷️ Model Name
ElimSoul-v1 – English to Malawian Sign Language Gloss Translator
🧾 Model Description
elimsoul-v1 is a custom-trained multilingual sequence-to-sequence language model designed to translate English phrases into Malawian Sign Language (MSL) gloss format. This model is built to bridge the educational gap for deaf and hard-of-hearing communities, especially in Malawi, through AI-powered sign language access.
The model is a fine-tuned version of facebook/mbart (or google/mt5-small, depending on your base) and focuses specifically on educational, spiritual, social, and conversational domains.
🌍 Use Cases
🏫 Educational settings: Help teachers communicate content in sign-based gloss.
🤝 Community tools: Used in ElimSoul app for student–teacher–parent sign communication.
🎬 Sign animation pipeline: Input for generating sign avatar animations via Blender.
📱 Mobile/online integration: Integrated in video chat & summarizer systems to display gloss output from other ElimSoul modules.
📦 Training Data
The model was trained on a manually curated dataset of over 1,000 English–Gloss pairs. Sources include:
Custom-written educational, spiritual, and behavioral sentences
Community-sourced gloss examples (e.g. “I want help” → “I WANT HELP”)
Real-world phrases for conversational AI in MSL
Expanded with tech, science, and faith-based short stories
Note: The glosses are based on MSL grammatical structure and may differ significantly from English word order.
🛠️ Model Architecture
Component Description
Base model mt5-small or facebook/mbart
Task type Sequence-to-sequence translation
Tokenizer SentencePiece / T5Tokenizer
Output format Uppercase GLOSS (e.g., YOU HELP ME)
🧪 Evaluation
Evaluation was done manually on:
✅ Gloss consistency
✅ Faithfulness to source meaning
✅ Handling of short vs long phrases
⚠️ Some limitations on very abstract input or uncommon expressions
📉 Limitations
Does not generate full sign video – only GLOSS format.
Short answers like "Yes", "No", and abstract sentences can be inconsistent if not present in the dataset.
Model currently supports only English input → MSL gloss output.
✅ Intended Use
✔️ For translation of short-to-medium educational texts
✔️ Use with ElimSoul gloss animation pipeline
✔️ Use in low-resource educational platforms
❌ Not intended for real-time sign recognition or ASL
🔐 License
✍️ Author and Maintainer
Juslin Waka (Juslin Jay)
Founder, ORTHO EduHub & ElimSoul
📍 Dzaleka Refugee Camp, Malawi
📧
waka2j@cmich.edu |
juslinwaka@gmail.com
🌐 ElimSoul Project
🙌 Acknowledgments
Special thanks to:
Deaf learners and educators from Malawi
MSL native users who reviewed gloss consistency
Open-source contributors and the Hugging Face ecosystem
