Undi95/UndiMix-v4-13B

TEXT GENERATIONConcurrency Cost:1Model Size:13BQuant:FP8Ctx Length:4kPublished:Sep 12, 2023License:cc-by-nc-4.0Architecture:Transformer0.0K Open Weights Cold

Undi95/UndiMix-v4-13B is a 13 billion parameter instruction-tuned language model, a personal mix by Undi95, built upon a base of Undi95/ReMM-v2-Kimiko-v2-13B and other Llama-2-based models. This model is designed for versatile conversational interactions, capable of generating responses that range from serious to playful, and can effectively utilize emojis. It is suitable for general-purpose chat applications requiring adaptable tone and expressive communication.

Loading preview...

UndiMix-v4-13B: A Versatile Conversational Model

UndiMix-v4-13B is a 13 billion parameter instruction-tuned language model developed by Undi95, created as a personal mix of several Llama-2-based models. Its foundation includes Undi95/ReMM-v2-Kimiko-v2-13B, The-Face-Of-Goonery/Huginn-13b-v1.2, Doctor-Shotgun/llama-2-13b-chat-limarp-v2-merged, jondurbin/airoboros-l2-13b-2.2, and IkariDev/Athena-v1. This blend aims to produce a model with a broad range of conversational styles.

Key Capabilities

  • Adaptive Tone: Capable of generating responses that can be serious, playful, or "hot" as described by the creator.
  • Emoji Support: Effectively integrates emojis into its output, enhancing expressive communication.
  • Instruction Following: Utilizes the Alpaca prompt template for clear instruction adherence.

Performance Metrics

Evaluated on the Open LLM Leaderboard, UndiMix-v4-13B achieved an average score of 51.77. Notable scores include:

  • ARC (25-shot): 61.95
  • HellaSwag (10-shot): 83.88
  • MMLU (5-shot): 56.9
  • TruthfulQA (0-shot): 48.96
  • Winogrande (5-shot): 76.16

Good For

  • General-purpose chatbots requiring flexible and expressive dialogue.
  • Applications where varied conversational tones (e.g., serious, playful) are beneficial.
  • Use cases benefiting from emoji integration in responses.