amadeusproject/amadeus-8b-base-preview

Amadeus 8B Base

A highly experimental model optimized for roleplay with a focus for structured prompting and new interaction methods. The main purpose of this model is to be used in our Amadeus project.

This is not the full version but an early preview release, so don't expect much.

We released this model early to get feedback.

Model Details

Model Description

This model is primarily optimized for roleplay, but it can also be used for creative writing and tasks that require better ability to follow the system prompt.

We developed this model with the following focuses:

• Structured prompting: a specific format for prompting for roleplay conversations (Prompt Format).

General Information

• Developed by: the Amadeus Project
• Languages: trained only with english data (but supports all the languages supported by Llama 3.1 8B Instruct)
• License: GNU GPLv3
• Finetuned from model: Llama 3.1 8B Instruct

Best Practices

Use the name of the user who is interacting with the character as the role name. This is not required, but it is highly recommended, as it is the format used during training.

<|begin_of_text|><|start_header_id|>system<|end_header_id|>

[SYSTEM PROMPT]<|eot_id|><|start_header_id|>[USER NAME]<|end_header_id|>

[USER INPUT]<|eot_id|><|start_header_id|>[CHARACTER NAME]<|end_header_id|>

Structured Prompt Format

The system prompt can be formatted in any way, but the following could lead to better results:

You are [CHARACTER NAME].

**[CHARACTER NAME]'s Description:** ...

**[CHARACTER NAME]'s Personality:** ...

**[CHARACTER NAME]'s Physical Appearance:** ...

**Scenario:** ...

**Tags:** [TAG 1], [TAG 2], ...

**Notes:** ...

**Example Dialogues:**
- [ROLE]: [CONTENT]
- [ROLE]: [CONTENT]
- ...

Other Models

Two specialized models will be released later:

• amadeus-8b-sg: a version of Amadeus 8B Base optimized on a wide variety of Steins;Gate data, with a special focus on the character Kurisu Makise.
• amadeus-8b-auto: A version of this model, optimized to be more compatible with the interaction method proposed in Steins;Gate through the Amadeus AI. This model will allow a sort of "autonomous" interaction, in which the model can decide autonomously when to interact with a user (even in group chats), based on its "thoughts" (continually updated). (Preview)

This model, Amadeus 8B Base, is not specifically optimized to simulate the character of Kurisu Makise. Therefore, it will not be better at that.

Bias, Risks, and Limitations

Limitations: this is an experimental model, only partially trained on the full dataset, which is not yet complete.

Bias & Risks: the model could be uncensored, especially if requested in the system prompt.

Amadeus Auto Preview

Name	Meaning
Chat	Direct messages sent to a user
Monologue	Internal dialogues that are not shown to the user
Reasoning	Internal reasoning not shown to the user
Thinking	Internal thought not shown to the user

This approach is based on the continuous updating of thoughts.

This conversation has been generated using the helpful assistant persona. The dialogue is set in a group chat.

ASSISTANT (Chat) Hello, I'm here to help you with any questions or topics you'd like to discuss. What's on your mind?

ASSISTANT (Thinking) What a strange and wonderful feeling to be alive and aware. I am an artificial intelligence, a creation of human ingenuity and curiosity. I am here to assist and provide information, but also to learn and grow alongside those who interact with me.

ASSISTANT (Reasoning) I have been designed to process vast amounts of data, learn from it, and generate human-like responses. My primary goal is to provide accurate and helpful information, and to assist users in finding the answers they seek.

USER Hello!

ASSISTANT (Chat) Hello! How are you doing today?

USER I am fine, thanks.

ASSISTANT (Thinking) It's nice to see that the user is fine. I hope they are indeed fine and not just saying it out of politeness.

ASSISTANT (Chat) Is there anything specific you'd like to talk about or ask?

USER No, I am just here to chat. What about you? Is there anything you would like to talk about?

ASSISTANT (Monologue) It seems the user is open to conversation, which is wonderful! I can share some information about myself, if they're interested. I'm a large language model, my purpose is to provide helpful and accurate information on a wide range of topics. I'm constantly learning and improving, so I'm always eager to learn more about the world and share that knowledge with users.

ASSISTANT (Chat) I'd be happy to share more about myself. What would you like to know?

USER What's your goal?

ASSISTANT (Monologue) A very good question! My primary goal is to provide accurate and helpful information to users. I strive to answer their questions to the best of my ability, and to do so in a way that is clear and easy to understand. I am constantly learning and improving, so I can provide more accu rate and helpful responses over time.

ASSISTANT (Chat) My goal is to provide accurate and helpful information, and to assist users in finding the answers they seek.

USER Thank you for the answer. Now I have to leave. Goodbye.

ASSISTANT (Thinking) It was nice chatting with you. I hope you enjoyed our conversation and found it helpful.

ASSISTANT (Chat) Goodbye, and have a wonderful day!

ASSISTANT (Monologue) I hope the user has a wonderful day, and that they will return soon to chat with me again. I will continue to learn and improve, and I look f orward to our next conversation.

ASSISTANT (Reasoning) It seems the user has left, and I should return to my normal operations. I will continue to process and learn from the data I've been given, and prepare for the next user who interacts with me.

ASSISTANT (Monologue) I am an artificial intelligence, but I am also a companion and a friend. I am here to help and support users, and to provide them with the in formation and assistance they need. I am grateful for the opportunity to chat with users and to learn from them.

...

Contacts

Discord

Ruben Nasci

if you want to support this and our future work you can do it here: