Spaces:
Sleeping
Sleeping
| title: Diabetes Assistant - Multilingual! | |
| emoji: 🔥 | |
| colorFrom: indigo | |
| colorTo: green | |
| sdk: gradio | |
| sdk_version: 5.33.0 | |
| app_file: app.py | |
| pinned: false | |
| license: cc-by-4.0 | |
| short_description: Multi-lingual Diabetes chatbot. Responses in text and audio. | |
| # Project Title: Diabetes Assistant | |
|  | |
| ## Objective | |
| The objective of this project was to showcase our individual learnings about large language models, translation application, chatbot, gradio and hugging face. | |
| ## Sources | |
| - ChatGPT | |
| - Copilot | |
| - Hugging Face | |
| - Gradio | |
| - OpenAI Whisper (https://openai.com/research/whisper) | |
| - Langchain (https://www.langchain.com/) | |
| - Amazon Polly (https://docs.aws.amazon.com/polly/latest/dg/what-is.html) | |
| - Helsinki-NLP/opus-mt models (https://huggingface.co/Helsinki-NLP) | |
| ## Citations | |
| This project utilizes models from the OPUS-MT project. We thank Jörg Tiedemann and Santhosh Thottingal for their work: | |
| - Tiedemann, J., & Thottingal, S. (2020). OPUS-MT – Building open translation services for the World. In Proceedings of the 22nd Annual Conference of the European Association for Machine Translation (pp. 479–480). European Association for Machine Translation. [https://aclanthology.org/2020.eamt-1.61](https://aclanthology.org/2020.eamt-1.61) | |
| - Tiedemann, J. (2020). The Tatoeba Translation Challenge – Realistic Data Sets for Low Resource and Multilingual MT. In Proceedings of the Fifth Conference on Machine Translation (pp. 1174–1182). Association for Computational Linguistics. [https://aclanthology.org/2020.wmt-1.139](https://aclanthology.org/2020.wmt-1.139) | |
| ## Method | |
| L3-AI Created an assistant to ask your diabetes questions and when needed translate responses to an alternate language. | |
| 1. <b>Transcription:</b> Individuals could either voice their questions by hitting the microphone, upload an mp3 of their question, or write their diabetes | |
| related questions within the Hugging Face Application. For questions that were either voice activated or mp3 uploaded we used <i>openai/whisper-large</i> to | |
| transcribe the audio into written format. | |
| 2. <b>LLM Model:</b> Using <i>WikipediaLoader</i>, we created a large language model that tapped into Wikipedia specifically grabbing information related to the diabetes | |
| question. | |
| 3. <b>Chatbot Response and Voice Over:</b> L3-AI added a feature that allowed our Hugging Face Application to verbalize the response from the LLM as well as provide responses in | |
| written format. We used <i>Amazon Polly</i>, to provide written text to speech. | |
| 4. <b>Translation:</b> <i>Helsinki-NLP</i> was used to translate the information provided from the LLM. | |
| 5. <b>Gradio:</b> L3-AI used the gradio application to organize and produce each level and response of the four different models utilized. | |
| 6. <b>Hugging Face:</b> Finally, L3-AI pushed all information to Hugging Face Application for speed as well as production. | |
| ## Interface | |
| https://huggingface.co/spaces/L3-AI/diabetes_assistant | |
|  | |
| ## Learnings | |
| Natural Language Processing (NLP): | |
| - Gained insights into NLP techniques and methodologies used for building our conversational agent. | |
| - Learned about tokenization, language modeling, and how to improve speed within our chatbot development. | |
| Model Selection and Evaluation: | |
| - Evaluate different language models such as LLM and Polly for their performance in generating human-like responses. | |
| - Compare model capabilities, including coherence, fluency, and ability to stay on topic. | |
| - Understand the strengths and limitations of each model in different conversational contexts. | |
| Fine-tuning: | |
| - Address issues such as speed and translation accuracy by fine-tuning model parameters and configurations. | |
| - Implement strategies to mitigate challenges such as text truncation and limited language support to enhance overall user experience. | |
| - Iterate on model architecture, hyperparameters, and data preprocessing techniques to achieve desired outcomes and user satisfaction. | |
| Hugging Face: | |
| - Emphasize the necessity of creating a comprehensive requirements document outlining dependencies, libraries, and configurations required for Hugging Face model integration. | |
| - Avoid reliance on Jupyter notebooks for production-level deployment due to limitations in scalability, version control, and reproducibility. | |
| Streamlit VS Gradio: | |
| - Recognized Streamlit's appeal for deployment purposes, particularly for its visually appealing characteristics and user interface elements. | |
| - However, prioritized Gradio for deployment due to its compatibility with the core functionality and focus of our model, prioritizing model performance and functionality over visualization aesthetics. | |
| ## Opportunities and Next Steps | |
| For L3-AI concept design we centered on diabetes however, we thought in future endeavors expanding to other disease states would enhance the work that was started. | |
| The source material limited which sources we could pull from due to API restrictions. | |
| ## Credits | |
| We would like to thank our pets who kept us company as we worked on coding and this application. |