Model Overview

l3utterfly/mistral-7b-v0.1-layla-v4 is a 7 billion parameter language model built upon the Mistral 7B architecture. Developed by l3utterfly and funded by Layla Network, this model is specifically fine-tuned for enhanced multi-turn conversation and character impersonation capabilities.

Key Characteristics

• Fine-tuning Dataset: Optimized using a pre-processed version of the OpenHermes 2.5 dataset, which involved removing refusals and AI assistant mentions, and splitting multi-turn dialogues. It also incorporates NSFW conversations from the Teatime dataset.
• Primary Application: Serves as the foundational model for Layla, an offline personal assistant, highlighting its suitability for interactive AI applications.
• Context Length: Supports an 8192-token context window, enabling more extensive and coherent conversations.

Performance Benchmarks

Evaluations on the Open LLM Leaderboard indicate a strong average performance across various reasoning and language understanding tasks:

• Avg.: 64.69
• AI2 Reasoning Challenge (25-Shot): 62.29
• HellaSwag (10-Shot): 83.36
• MMLU (5-Shot): 64.32
• TruthfulQA (0-shot): 43.14
• Winogrande (5-shot): 79.56
• GSM8k (5-shot): 55.50

Use Cases

This model is particularly well-suited for:

• Multi-turn conversational agents: Excelling in maintaining context and coherence over extended dialogues.
• Character impersonation: Capable of adopting specific personas for role-playing or interactive storytelling.
• Offline personal assistants: Designed to function effectively in environments without constant internet connectivity, as demonstrated by its use in Layla.