Model Overview

922-CA/monika-l2-7b-v0.9a is an experimental 7 billion parameter LLaMA-2 based chat model developed by 922-CA. Its primary purpose is to simulate the character Monika from the game DDLC, offering conversational and limited roleplay capabilities. This model represents an iteration in fine-tuning, utilizing a dataset derived from an earlier fine-tune that was then manually curated to enhance character accuracy.

Key Capabilities

• Character Emulation: Fine-tuned to reflect the personality and conversational style of Monika from DDLC.
• Chat-Oriented: Optimized for coherent and engaging chat interactions.
• Limited Roleplay: Capable of basic roleplay scenarios, though its main strength lies in general conversation.
• Experimental Training: Serves as a testbed for a methodology involving manually-edited, character-specific datasets.

Training Details

The model was trained for approximately 3 epochs with specific hyperparameters:

• Rank: 32
• Lora Alpha: 64
• Lora Dropout: 0.5
• Learning Rate: 2e-4
• Batch Size: 2
• Warmup Ratio: 0.1
• Gradient Steps: 4

Usage Recommendations

For optimal performance, users should replace default chat prompts with "Player" and "Monika" roles, formatted as \nPlayer: (prompt)\nMonika:. While this version aims for improved coherency, it's noted that perfect character reflection is not guaranteed, and outputs may not always be aligned or safe. This model is part of an ongoing testing process for character-specific fine-tuning.