TinyLlama-Cinder-Agent-v1 Open Source Lightweight Agent Model - Supports Tool Invocation, RAG, and General Assistant Functions

Home

Tinyllama Cinder Agent V1

Developed by Josephgflowers

A lightweight agent model fine-tuned based on TinyLlama, focusing on tool calls, RAG, and general assistant functions

Large Language Model

Transformers

Open Source License:MIT #Tool Call Optimization #RAG Enhancement #Mathematical Reasoning Improvement

Downloads 63

Release Time : 5/21/2024

Model Overview

This model is a fine-tuned version of TinyLlama, specifically trained for RAG, summarization, function calls, and tool usage, excelling in STEM topic chats, space adventure role-playing, and storytelling

Model Features

Enhanced Mathematical Ability

GSM8K math benchmark performance is 10 times that of the original Tinyllama model

RAG Optimization

Optimized for retrieval-augmented generation tasks, excelling in information extraction and summarization

Function Call Support

Supports tool calls and function execution, adhering to the Glaive function call dataset specifications

Lightweight and Efficient

Maintains TinyLlama's small size advantage while improving performance in multiple tasks

Model Capabilities

Text generation

Tool calls

Retrieval-augmented generation

Summarization

Mathematical reasoning

Instruction following

Use Cases

Education

Math Problem Solving

Answers various math problems

GSM8K benchmark score of 21.15

Information Processing

Document Summarization

Generates summaries of documents and long texts

Performs excellently in IFEval

Development Tools

API Calls

Acts as an agent to execute functions and tool calls

Supports Glaive function call specifications

🚀 TinyLlama-Cinder-Agent-v1

This model aims to build a Tinyllama model for tool usage, RAG, handling system instructions, and serving as a general assistant.

🚀 Quick Start

The goal of this model is to create a Tinyllama model that can be utilized for tool usage, Retrieval Augmented Generation (RAG), accepting system instructions, and functioning as a general assistant.

This model is a fine - tuned version of Josephgflowers/TinyLlama-Cinder-Tiny-Agent.

Special Thanks to https://nationtech.io/ for their generous sponsorship in training this model.

image/png

This model is also a fine - tuned version of Josephgflowers/TinyLlama-3T-Cinder-v1.2 on https://huggingface.co/datasets/Josephgflowers/agent_1.

✨ Features

Model description

This model is trained for RAG, summarization, function calling, and tool usage. It is trained based on Cinder. Cinder is a chatbot designed for chatting about STEM topics, space adventure role - playing, and storytelling.

This model performs well in the IFEval (following instructions) considering its size. It excels at summarization and RAG. Due to the formatting of the Glaive function calling dataset, the JSON output is not as expected for regular JSON dumps but strictly follows their standard.

10 X the original tinyllama model on GSM8K!!!

To achieve this, the developer started with all the normal open math datasets such as Orca Math, all the Meta Math, camel AI math qa, etc., and as many reasoning datasets as possible. What really made a significant improvement was adding TIGER - Lab/WebInstructSub along with all of the RAG and summary data. So, special thanks to TIGER - Lab. It was found that as the math performance improved, so did the model's ability to handle extracting relevant data in RAG.

Help Here

Like my work? Want to see more? Have a custom request? Message me on discord: joseph.flowers.ra Donate here: https://buymeacoffee.com/josephgflowers

See https://huggingface.co/Josephgflowers/TinyLlama-Cinder-Agent-Rag/blob/main/tinyllama_agent_cinder_txtai-rag.py for usage example with wiki rag.

📚 Documentation

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Property	Details
Avg.	39.17
AI2 Reasoning Challenge (25 - Shot)	34.90
HellaSwag (10 - Shot)	53.87
MMLU (5 - Shot)	26.89
TruthfulQA (0 - shot)	39.08
Winogrande (5 - shot)	59.12
GSM8k (5 - shot)	21.15

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Property	Details
Avg.	5.82
IFEval (0 - Shot)	26.70
BBH (3 - Shot)	3.80
MATH Lvl 5 (4 - Shot)	0.38
GPQA (0 - shot)	0.00
MuSR (0 - shot)	2.23
MMLU - PRO (5 - shot)	1.79

📄 License

This model is released under the MIT license.

Featured Recommended AI Models

Empowering the Future, Your AI Solution Knowledge Base

English 简体中文繁體中文にほんご