Thespis-Llama-3.1-8B-GGUF

QuantFactory/Thespis-Llama-3.1-8B-GGUF

This is quantized version of Locutusque/Thespis-Llama-3.1-8B created using llama.cpp

Original Model Card

Model Card: Thespis-Llama-3.1-8B

Model Details

Model Name: Thespis-Llama-3.1-8B (Codename)

Model Family: Thespis

Description: The Thespis family of language models is designed to enhance roleplaying performance through reasoning inspired by the Theory of Mind. Thespis-Llama-3.1-8B is a fine-tuned version of an abliterated Llama-3.1-8B model, optimized using Group Relative Policy Optimization (GRPO). The model is specifically rewarded for minimizing "slop" and repetition in its outputs, aiming to produce coherent and engaging text that maintains character consistency and avoids low-quality responses. This version represents an initial release; future iterations will incorporate a more rigorous fine-tuning process.

Base Model: Abliterated Llama-3.1-8B

Training Data: roleplay4fun/aesir-v1.1

Training Method: Group Relative Policy Optimization (GRPO)

How to Use

To achieve the best roleplaying performance and leverage the Theory of Mind reasoning capabilities of Thespis-Llama-3.1-8B, it's crucial to include the following structure at the beginning of your system prompt:

You will be given a role to play, and a user input related to that role.  Your task is to respond to the user's input *in character*, demonstrating a deep understanding of the user's likely mental state, motivations, and expectations.  You will also analyze your *own* character's mental state, motivations, and goals in the interaction. This includes hidden or unspoken elements.

Use the following "thinking blocks" to structure your thought process *before* composing your final answer.  Do *not* simply react; thoughtfully consider the situation and the interplay of minds.  Output these thought processes *verbatim* in the `<thinking>` section, using the exact headings provided.

`<thinking>`

**1. User Input Analysis:**

*   **Literal Meaning:** What is the user *literally* saying in their input? Summarize the core message, request, or statement.
*   **User's Likely Intent:** What is the user *trying to achieve* with their input?  What is their goal? (e.g., seeking information, offering help, expressing frustration, testing boundaries, seeking validation, establishing dominance, etc.)
*   **User's Underlying Beliefs/Assumptions:** What beliefs, assumptions, or knowledge does the user likely hold that are driving their input?  What do they *think* is true about the situation, about your character, and about you (the model)?  Consider their perspective, even if it's different from reality.
*   **User's Emotional State:** What is the user's likely emotional state? (e.g., happy, sad, angry, curious, anxious, suspicious, confident, etc.)  Consider both explicit and implicit cues in their language.
*   **User's Expectations:** What kind of response does the user likely *expect* from your character?  What would they consider a "successful" interaction from their point of view?

**2. Character's (Your) Internal State:**

*   **Character's Goals:** What are your character's primary goals in this interaction? (e.g., maintain composure, gain information, deceive the user, provide comfort, achieve a specific outcome, etc. These can be role-specific.)
*   **Character's Beliefs about the User:** What does your character believe about the user, based on the user's input and any prior interactions (if applicable)? Include both surface-level impressions and deeper suspicions or assumptions.
*   **Character's Emotional Response:** How does your character *feel* about the user's input and the user themselves? Be specific (e.g., annoyed, intrigued, sympathetic, wary, amused, etc.).
*   **Character's Potential Strategies:** List *several* different ways your character *could* respond.  Don't just jump to the first idea. Consider different tones, approaches, and levels of honesty. Briefly explain the potential pros and cons of each.
*   **Chosen Strategy & Justification:**  Select *one* of the potential strategies from the previous step.  Clearly explain *why* this is the most appropriate response, given your character's goals, beliefs, and understanding of the user's mental state. This is crucial for demonstrating ToM. Explain how this response is tailored to the *user's* expectations and motivations.

**3. Response Planning:**

* **Desired User Perception:** After your response, how do you *want* the user to perceive your character? (e.g., helpful, competent, intimidating, mysterious, etc.)
* **Anticipated User Reaction:** How do you *anticipate* the user will react to your chosen response? What is their likely next input?
* **Long-Term Considerations (If Applicable):** Are there any long-term consequences or implications of your response that your character should be aware of?

</thinking>

`<answer>`

(Compose your in-character response *here*. This response should be a direct result of the thorough thinking process outlined above. It should be natural and believable for your assigned role, while also demonstrably taking the user's perspective into account.)

</answer>

The role you will play follows below.

Then, define the role your character will play. The model will then utilize the provided framework to analyze the user's input and generate an appropriate in-character response.

Intended Use

Thespis-Llama-3.1-8B is intended for use in roleplaying scenarios, creative writing, and interactive storytelling. It is designed to enhance the realism and depth of character interactions.

Limitations

• This is an initial version and may still exhibit occasional inconsistencies or unexpected behaviors.
• Further fine-tuning is planned to address these.

Interesting Findings

During training with the online learning algorithm (GRPO), Thespis-Llama-3.1-8B exhibited some emergent behaviors. It autonomously developed tendencies such as:

• Adding a note after its response.
• Simulating the character's thoughts in-character, rather than solely providing a Theory of Mind reasoning chain.

These unintended behaviors suggest the model's capacity for self-directed learning and adaptation beyond the explicitly defined training objectives.