Gpt-Roleplay: A Fine-Tuned Persona Model

🚀 Model Overview

This model is a specialized GPT-2 variant designed for deep, immersive roleplay. It has undergone a dual-stage training process to transition from a technical assistant to a creative narrator.

🛠 The Training Journey

Phase 1: The Technical Foundation

Initially, the model was fine-tuned on a 10,000-row technical Q&A dataset. This established a robust baseline for dialogue structures and logical consistency, but resulted in a heavy bias toward 'instructional' and 'professional' language.

Phase 2: The Roleplay Pivot

To break the technical habit, we introduced a 10,000-row 'Pure Roleplay' dataset. This dataset featured complex scenarios ranging from high-fantasy mountains to sci-fi spaceports, focusing on atmospheric descriptions and character-driven dialogue.

Phase 3: The 'Persona Guard' System

We observed 'technical leakage' where the model would generate meta-commentary (e.g., 'provide visuals for analysis'). To fix this, we implemented:

• Aggressive Repetition Penalties (2.2 - 3.0): Forces the model away from its pre-trained technical loops.
• Token-Level Blocking: Using bad_words_ids to prevent the generation of words like 'metrics', 'documentation', and 'workshop'.
• Persona Injection: Framing every prompt with a strong identity (e.g., 'A storyteller from a forgotten age').

🎮 Recommended Inference Settings

For best results, use these parameters in your generation config:

• Temperature: 1.15
• Repetition Penalty: 2.5
• Top-K: 40
• Top-P: 0.9

⚠️ Known Limitations

Due to its GPT-2 architecture, the model can sometimes output NATO phonetic noise (e.g., 'Whiskey Oscar'). Our generate_roleplay function includes a regex-based blocklist to strip this noise automatically.