Josephgflowers/TinyLlama-Cinder-Agent-v1

Hugging Face
TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:1.1BQuant:BF16Context Size:2kPublished:May 21, 2024License:mitArchitecture:Transformer0.0K Open Weights Featherless Exclusive Warm

Josephgflowers/TinyLlama-Cinder-Agent-v1 is a 1.1 billion parameter language model, fine-tuned from Josephgflowers/TinyLlama-3T-Cinder-v1.2, with a 2048 token context length. It is specifically optimized for tool usage, RAG (Retrieval Augmented Generation), function calling, and following system instructions. This model demonstrates strong performance in summary tasks and instruction following (IFEval) for its size, and shows significantly improved mathematical reasoning compared to the original TinyLlama.

Loading preview...

Josephgflowers/TinyLlama-Cinder-Agent-v1: A Compact Agentic LLM

Josephgflowers/TinyLlama-Cinder-Agent-v1 is a 1.1 billion parameter language model, fine-tuned from Josephgflowers/TinyLlama-3T-Cinder-v1.2, designed for agentic applications. This model builds upon the Cinder chatbot's foundation, which focuses on STEM topics, space adventure role-playing, and storytelling, enhancing its capabilities for structured tasks.

Key Capabilities

  • Tool Usage & Function Calling: Optimized for integrating with external tools and executing functions based on prompts, though its JSON output format adheres strictly to the Glaive function calling dataset standard.
  • Retrieval Augmented Generation (RAG): Excels at extracting and utilizing relevant information for generation tasks, demonstrating strong performance in RAG scenarios.
  • Instruction Following (IFEval): Shows good ability to follow instructions for its size, making it suitable for tasks requiring precise adherence to commands.
  • Summary Generation: Proficient in summarizing text, a core component for many agentic workflows.
  • Enhanced Mathematical Reasoning: Achieves a significant 10x improvement on GSM8K compared to the original TinyLlama, benefiting from extensive training on math and reasoning datasets like Orca Math, Meta Math, and camel AI math qa, supplemented by TIGER-Lab/WebInstructSub.

Good For

  • Resource-constrained environments: Its small size (1.1B parameters) makes it efficient for deployment where computational resources are limited.
  • Agentic workflows: Ideal for applications requiring tool use, function calling, and precise instruction following.
  • RAG-based systems: Effective for tasks that involve retrieving and synthesizing information from external knowledge bases.
  • Educational and STEM-focused applications: Benefits from its Cinder heritage and enhanced mathematical reasoning, making it suitable for STEM-related chat and problem-solving.
  • Rapid prototyping: Its compact nature allows for quicker experimentation and iteration in developing agent-based solutions.

Finetunes
4 models