PocketDoc/Dans-PersonalityEngine-V1.1.0-12b

Dans-PersonalityEngine-V1.1.0-12b Overview

Dans-PersonalityEngine-V1.1.0-12b is a 12 billion parameter language model developed by PocketDoc, based on the mistralai/Mistral-Nemo-Base-2407 architecture. It boasts an extensive context length of 32768 tokens, enabling it to handle complex and lengthy interactions. The model is full finetuned for 2 epochs, totaling 88 hours of training on a single H200 SXM.

Key Capabilities

• Multifarious Task Handling: Proficient in co-writing, roleplay, sentiment analysis, and summarization.
• Diverse Training: Trained on a wide array of one-shot and multi-turn instructions, tool use, role-playing scenarios, and text adventure games.
• Specialized Datasets: Incorporates unique datasets for memory, military/survival knowledge, mathematics (including Chain of Thought reasoning), programming (LeetCode, code feedback), and creative writing (co-writer, interactive fiction).
• System & Tool Usage: Features training for agent behavior, shell commands, and API/function calling.
• Optimized for Benchmarking: Includes specific datasets (Dans-Benchmaxx, Dans-Benchmaxx-COT) to prepare the model for "answer only" and Chain of Thought benchmarks, preventing excessive verbosity when direct answers are required.
• ChatML Format: Utilizes the standard ChatML prompting format for easy integration.

Good For

• Interactive Applications: Ideal for chatbots, virtual assistants, and role-playing games due to its strong co-writing and personality development capabilities.
• Content Generation: Suitable for creative writing, story generation, and text editing tasks.
• Technical Assistance: Can assist with basic programming explanations and debugging, as well as complex mathematical problem-solving.
• Data Processing Pipelines: Effective for sentiment analysis, summarization, and data preparation within larger systems.
• Instruction Following: Highly capable of adhering to system messages and complex multi-turn instructions.