Hjgugugjhuhjggg/mergekit-ties-tzamfyy

Hugging Face
TEXT GENERATIONConcurrent Unit Cost:1Model Size:1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Nov 22, 2024Architecture:Transformer Featherless Exclusive Warm

Hjgugugjhuhjggg/mergekit-ties-tzamfyy is a 1 billion parameter language model created by Hjgugugjhuhjggg using the TIES merge method, based on meta-llama/Llama-3.2-1B. This model integrates capabilities from multiple Llama-3.2-1B variants, including those fine-tuned for mathematical reasoning (MGSM8K, GSM8K), medical question answering (medmcqa), common sense QA, SQL, and code generation. It is designed to offer a broad range of specialized functionalities within a single compact model.

Loading preview...

Model Overview

Hjgugugjhuhjggg/mergekit-ties-tzamfyy is a merged language model built upon the meta-llama/Llama-3.2-1B base, utilizing the TIES merge method. This approach combines the strengths of multiple specialized Llama-3.2-1B variants into a single, cohesive model.

Key Capabilities

This model integrates diverse functionalities from its constituent models, including:

  • Mathematical Reasoning: Incorporates models fine-tuned on datasets like MGSM8K and GSM8K, enhancing its ability to handle mathematical problems.
  • Question Answering: Includes components optimized for medical question answering (medmcqa) and common sense QA.
  • Code and SQL Generation: Merges models with capabilities in SQL and general code tasks.
  • Multilingual Support: Features a component fine-tuned on Hinglish medical datasets, suggesting some multilingual adaptability.

Merge Details

The merge process involved a variety of Llama-3.2-1B instruction-tuned and pre-trained models, each contributing specific expertise. The TIES method, with a density of 0.5, was used to combine these models, aiming for a balanced integration of their respective strengths. This allows the model to perform well across a spectrum of tasks that its individual components were specialized for.