PetrolLM: A Fine-tuned Model for Creative Writing and Roleplay
PetrolLM is a 7 billion parameter model based on Mistral-7B-v0.1, specifically fine-tuned for creative writing and roleplay scenarios. It utilizes QLoRA (4-bit precision) for efficient adaptation.
Key Capabilities and Features
- Optimized for Creative Text Generation: Excels at producing descriptive, immersive, and engaging text for roleplay and narrative applications.
- Diverse Training Data: Fine-tuned on a curated dataset of approximately 5800 samples, including:
- AICG Logs
- PygmalionAI/PIPPA
- Squish42/bluemoon-fandom-1-1-rp-cleaned
- OpenLeecher/Teatime
- Norquinal/claude_multiround_chat_1k
- jundurbin/airoboros-gpt4-1.4
- totally-not-an-llm/EverythingLM-data-V2-sharegpt
- GPT-Enhanced Dataset: Training samples were back-filled or converted using GPT-4/GPT-3.5-turbo-16k to fit a specific prompt format.
- Specialized Prompt Format: Designed to work effectively with a prompt structure similar to the original SuperHOT prototype, facilitating structured roleplay interactions.
Recommended Use Cases
- Interactive Fiction and Roleplay: Ideal for generating character dialogue, environmental descriptions, and narrative progression in text-based games or roleplay platforms.
- Creative Content Generation: Suitable for writers looking for assistance in generating descriptive passages or exploring different narrative styles.
Technical Details
- Base Model: Mistral-7B-v0.1
- Fine-tuning Method: QLoRA (4-bit precision)
- LoRA Parameters: Rank 64, Alpha 16, Dropout 0.1
- Training: BF16, 2 epochs, Cutoff Length 2048
For optimal performance in UIs like SillyTavern, specific Last Output Sequence prompts are suggested to encourage detailed and multi-paragraph responses.