Week #4 #

Progress Report #

External Feedback #

During this week, our team shared the prototype of the project with our pears to get feedback on the basic functionality of MVP. Since the machine learning model is still in development, and the prototype only has neutral language mode, the feedback was mostly focused on the user interface of Telegram bot and the ability to interact with it.
The main points of feedback and suggestions for improvement were:

The user interface needs to be more intuitive and user-friendly. The user should be able to understand the purpose of the bot and how to interact with it without any prerequisite knowledge.
Even though the neutral language model is able to hold a conversation, the main functionality of the bot to mimic someone’s speech is not yet implemented. This feature is crucial for the project and should be implemented as soon as possible.
The design of the website is simple and clean with a clear call to action. However, it lacks the intended functionality as it is not yet connected to the other parts of the project.

Testing #

Since the main priority of this week was to connect the Telegram bot with the machine learning model, the focus of testing was on the model training and its interaction with the user. The testing was done by interacting with the bot and checking the responses of the model. The main points of testing were:

The bot should be able to mimic the language style of the user from the training data.
The bot should be able to hold a conversation with the user and respond to their messages.

The testing results showed that the model has problems that occur after training on the dataset. The model is not able to generate coherent responses and mimics the language style of the user. This is a critical issue that needs to be addressed before the next iteration. The website and the Telegram bot were not unit-tested as the development and design choices are yet to be finalized.

By collecting feedback and testing the prototype, our team identified the main areas that need to be refined. Our initial goal for the week was to advance the machine learning model to the point where it can mimic the language style of the user. However, due to the lack of practical knowledge in the field of fine-tuning language models, we could not achieve the desired results during this week. Therefore, after analyzing the feedback and testing results, we decided to focus on terminating the problems with the model and improving the user interface of the bot and the website in the next iteration.

Challenges & Solutions #

The main challenges were:

The machine learning model is not able to generate coherent responses and mimic the language style of the user. The problem is related to the degradation of the model’s performance after training on the dataset. To address this issue, our team will try following approaches:
- Fine-tuning the model on a smaller dataset to prevent overfitting.
- Experimenting with different hyperparameters and training strategies to improve the model’s performance.
- Experimenting with different language models and architectures to find the best fit for our project.
- Searching for the mistakes in code and data preprocessing that could lead to the model’s poor performance.
Difficulties in connecting the Telegram bot with the machine learning model. The problem was solved by diving deeper into the Telegram API documentation and learning how to properly request the model’s responses.

Conclusions & Next Steps #

This week was challenging for our team as we faced difficulties in advancing the machine learning model to the desired level. However, the feedback and testing results helped us identify the main areas that need to be refined in the next iteration.
Next week our team will focus on dealing with the problems of the machine learning model and improving the user interface of the bot and the website. We plan to have a fine-tuned model that can mimic the language style of the user and hold a conversation via improved Telegram bot. Additionaly, we will finalize the design of the website and connect it to the backend part of the project to make it functional and usable.