Undi95/ReMM-L2-13B-PIPPA

Undi95/ReMM-L2-13B-PIPPA Overview

Undi95/ReMM-L2-13B-PIPPA is a 13 billion parameter language model that represents a recreation of the original MythoMax-L2-13b architecture. This model distinguishes itself through an updated merging process that incorporates several advanced Llama-2-based models and integrates the PIPPA dataset for enhanced performance.

Key Capabilities & Architecture

• Base Model: Built upon the robust TheBloke/Llama-2-13B-fp16 foundation.
• Merged Components: Integrates a diverse set of models including The-Face-Of-Goonery/Chronos-Beluga-v2-13bfp16, jondurbin/airoboros-l2-13b-2.1, NousResearch/Nous-Hermes-Llama2-13b, and The-Face-Of-Goonery/Huginn-13b-v1.2 through a multi-stage TIES-merging process.
• Dataset Integration: Further refined by merging with the zarakiquemparte/PIPPA-ShareGPT-Subset-QLora-13b dataset at a 0.18 weight, aiming to improve instruction following and conversational abilities.
• Prompt Format: Utilizes the Alpaca prompt template, making it compatible with a wide range of instruction-tuned applications.

Performance Benchmarks

Evaluated on the Open LLM Leaderboard, ReMM-L2-13B-PIPPA demonstrates a balanced performance across various tasks:

• Average Score: 52.58
• Reasoning (ARC): 59.73 (25-shot)
• Common Sense (HellaSwag): 83.12 (10-shot)
• General Knowledge (MMLU): 54.1 (5-shot)
• Truthfulness (TruthfulQA): 49.94 (0-shot)
• Winogrande: 74.51 (5-shot)
• Math (GSM8K): 2.96 (5-shot)
• Reading Comprehension (DROP): 43.69 (3-shot)

Use Cases

This model is suitable for general-purpose language tasks, including instruction following, conversational AI, and text generation, benefiting from its diverse merged components and the PIPPA dataset's influence.