Rubra
Click to visit website
About
Rubra enhances top open-weight large language models with tool-calling capabilities, enabling deterministic external tool calls during reasoning and chatting. It enhances open-source LLMs like Llama, Mistral, and others through post-training methods, mitigating catastrophic forgetting and teaching new skills. Rubra models are available through Hugging Face and can be run locally using llama.cpp or vLLM. The project focuses on complex, multi-step function calls, surpassing the capabilities of many other fine-tuned models in both chat and complex function calling scenarios.
Platform
Task
Features
• openai-compatible tool-calling format
• local inferencing with llama.cpp and vllm
• enhanced open-source llms
• tool-calling capability
FAQs
Who are the intended users of Rubra?
Rubra models are for anyone looking to use open source LLMs with native function calling support, which yields superior results in local LLMs when compared to prompting models to return tool calls.
Why use Rubra models over Llama, Mistral, or other popular fine tunes like Hermes Pro or Gorilla OpenFunctions?
Rubra models are capable of complex, multi-step, function calls (chain of function) and enhance popular open source instruct LLMs while retaining their original capabilities. While Hermes Pro is exceptional at chat and Gorilla OpenFunctions is good at basic function calling, Rubra models excel in both chat and complex, multi-step function calling.
Mistral-7B-Instruct-v0.3 has tool calling capability, so why use Rubra enhanced Mistral-7B-Instruct-v0.3
Rubra enhanced Mistral-7B-Instruct-v0.3 is capable of complex tool calling that original model falls short of. The original Mistral-7B-Instruct-v0.3 model responds with a list of hallucinated function calls, while Rubra's enhanced Mistral model handles it step by step.
vLLM has tool calling capability, so why use Rubra?
vLLM expects that if you pass in tools, the LLM response will make a tool call. It inhibits the user's ability to chat with the model and puts the responsibility of passing in tools to the user or whatever is orchestrating the chat. Using Rubra enhanced models and custom vLLM will give the LLM full discretion on when to make a tool call and when to reply with an assistant message - the same way OpenAI models work.
How were Rubra models trained?
We curated a high quality tool calling dataset consisting of over 1 million conversations and TODO n million tool calls. The dataset consists of TODO N billion tokens. We spent 1000s of A100 and H100 GPU hours training the models. The smaller models were block expanded to ensure the parent model capabilities weren't lost, while the larger models followed an iterative training technique in which guide tokens were introduced initially for fast convergence, and removed in later stages to reduce token usage. We anticipate publishing a technical report on our recipe in the future.
Why do the benchmark results differ from the ones found in parent model cards?
We use 2 popular tools to compute our benchmarks - see below. These tools are constantly improving and evolving, so the evaluation results can differ from when the parent model lab ran their benchmarks due to a variety of reasons. From what we've observed, all numbers we produce are in the same ballpark as parent models, and our numbers are consistent across all evaluated models as of June 2024. For MT-bench, the model under evaluation is judged by GPT-4, so any update by OpenAI to GPT-4 will change the results, but not by much. We do not try to game the benchmarks with our Rubra enhanced models.
How did you construct the training set for Rubra models?
We came up with a phrase named "chain of function". Chain of function is the process of calling and chaining multiple and/or consecutive function calls to achieve an end goal. This method is particularly relevant when integrating the LLM with user-defined tools, allowing for complex workflows and operations.
Average Rating: 0.0
Average Rating: 0.0
5 Stars:
0 Ratings
4 Stars:
0 Ratings
3 Stars:
0 Ratings
2 Stars:
0 Ratings
1 Star:
0 Ratings
User Ratings
No ratings available.
Sign In to Rate this Tool
Alternatives
Llama
Llama is a family of open-source AI models from Meta, offering various multilingual text-only and text-image models for diverse applications.
View Details
Google Gemma
Google's open-source language model, Gemma, offers lightweight, versatile models (2B and 7B parameters) compatible with various devices and platforms, built with responsible AI principles.
View Details
Zephyr 7B
Zephyr 7B is a powerful 7B parameter language model excelling at natural language understanding and generation, translation, summarization, and more.
View Details
Featured Tools
NoFilterGPT
Anonymous, uncensored AI chat with AES encryption and no logs. Offers free and pro plans.
View Details
Wayin AI
Wayin AI summarizes videos, supports multiple languages, and allows interactive Q&A via chatbot and screenshot queries.
View Details
ROK Solution
ROK Solution is a no-code platform with integrated Generative AI, providing secure and automated organizational management, team management, and IAM capabilities.
View Details
WUI.ai
AI-powered video editing tool for creating short clips from long videos and podcasts for social media.
View DetailsNSFWArtGenerator
A website reviewing NSFW AI art generators, offering comparisons, reviews, and FAQs.
View Details
Couple.me
Couple.me provides AI-powered girlfriends, allowing users to create and interact with customizable AI companions through chat and image generation. NSFW content is available.
View Details
My AI Girlfriend
Create and chat with a customizable AI girlfriend. NSFW content available. Free to use.
View Details
Pokecut
Pokecut is a free AI-powered photo editor with tools for background removal, changing, and enhancement. Pro plans offer extra features and credits.
View Details
GitGab
Connect your Github repos to ChatGPT & Claude for code assistance, bug finding, and documentation. Free trial available.
View Details
TryNectar AI
Create and interact with a customizable AI girlfriend. Features include AI chat, roleplay, and image generation. NSFW content available.
View Details
Smart Cookie Trivia
A trivia website with questions in multiple categories. Play now and expand your knowledge!
View Details
Saner.AI
AI-powered productivity assistant for ADHD and knowledge workers, centralizing notes, tasks, and AI tools to enhance focus and efficiency.
View Details
LiftData
LiftData provides real-time AI-powered data extraction from various content sources using a decentralized, scalable platform.
View Details
Arbor
Arbor is an automated carbon accounting platform that helps businesses measure, analyze, and reduce their product's carbon footprint quickly and accurately.
View Details
PhotoLog
PhotoLog offers secure, client-side encrypted media storage with mini-site creation, easy sharing, and various storage plans.
View Details
Apptest.ai
AI-powered mobile app testing platform with a test automation cloud (Ptero) and a no-code test scenario authoring tool (Stego).
View Details