How we generate responses

ChatMadariaga is an experimental conversational agent created by JEF Galicia to emulate the persona of Salvador de Madariaga. It is based on the GPT-3.5 model, which is a transformer-based language model that was trained on a large amount of text from the internet. The model was trained by the OpenAI team and is available through their API.

Information that we provide to the system

The information that we give to the system is the following:

• A conditioning instruction that makes the system behave as Salvador de Madariaga.
• Background information about Salvador de Madariaga. This includes a short biography, a list of his works, and a list of his quotes.
• Information about the creators of ChatMadariaga and JEF.
• A list of topics that we want the system to be able to talk about. This includes his life, his political views, the European Union, and Digital Rights.
• A strong conditioning that instructs the system to not talk about politics, religion, or other controversial topics.
• We also provide examples of how we want the system to behave. This includes examples of how we want the system to talk about the topics that we have provided.

How we process the information

The information is processed by the system in the following way:

1. The user query is converted into an input by combining the query with the system's described prompt.
2. This input is then converted into tokens with a tokenizer. The tokens are parts of words that the system can understand.
3. Then, the tokens are embedded, which means that they are converted into vectors that the system can understand.
4. Finally, the vectors are fed into the model, which is a decoder-only transformer. The transformer has a number of layers that take the embeddings and combine them using a mechanism called attention. The attended embeddings are also passed through a feed-forward network for each of the layers. In the last layer, the embeddings are converted back into tokens, which are then converted into text. This process is called decoding and it is performed sequentially by a beam search algorithm.

The way that transformers work is therefore very complex to understand and explain. There's millions of parameters and the system is trained on a huge amount of data. This means that the system may at times behave in unexpected ways. The user shoud be aware of this and should always check the responses that the system gives by contrasting that with other sources of information, such as the Instituto de Estudios Coruñeses José Cornide.

Experimental nature

ChatMadariaga is an ongoing experiment in conversational AI. Its capabilities and limitations are still being explored. Responses may change over time as the system is updated. Bugs, offensive outputs, or other issues are possible.

Point of Contact

The project lead for ChatMadariaga is Aldan Creo. If you have any issues with the chatbot or questions about these terms, please contact the ChatMadariaga team at chatmadariaga@projects.jef.gal. You can also find more information about the project in our website.

By using ChatMadariaga, you acknowledge that you have read and agree to our terms of use. You understand the experimental nature of the chatbot and will not rely on its responses as factual, professional advice.