I'm excited to share with you the first release of the X-box AI framework.
AKA the @xenopus_v1 framework
github.com/Nim-Network-Fo…
You can try out our demo right now and create your first fine-tuned model!
This represents the culmination of over 1.5 years of work by me and the team at @nim_network, where we've focused on building crypto-native agents and AI applications. I'm incredibly grateful to partner with @YoniCombinator and collaborate with the talented NIM team and @__OhadDahan__ from @blockmesh_xyz
I've also had the privilege to learn from brilliant minds in AI and witness SOTA developments before the current AI boom.
Before we dive into the details, a few notes and principles:
Fine-tuning represents one of AI's greatest challenges. As post-training emerges as a key milestone in AI advancement, it requires innovative approaches to data collection, preparation, and fine-tuning. It demands extensive collaboration across all technical layers. That's why we aim to build a movement powered by a growing community of partners, passionate AI engineers, users, and data providers. We welcome collaboration from anyone who shares our commitment to open knowledge.
No BS features: We're not building for the sake of announcements. We aim to keep the framework as lean as possible with isolated components. Instead of creating 50+ Discord, TikTok, and L2 integrations, we're focusing on one core challenge: creating leading practices for data collection and fine-tuning of highly customized agents and models. Every feature we release will be validated with in-production agents.
We are an extremely dynamic, lean, and aggressive team: We live by 0-1 sprints and plan to move fast on expanding the framework and ecosystem around it, delivering completely new primitives (more on that in the coming weeks).
Current features and structure of the framework
We're going through massive refactoring of our internal codebase to make it simple to use yet robust. The current repo will expand rapidly based on that.
Demo: We're starting with the Bully demo that guides you through creating your own customized bully model, from dataset creation to training (which runs easily and free with our Google Colab notebook). The demo section will expand with more end-to-end examples as we add new features.
X-models with PoFT (proof of fine-tuning): This section indexes customized models that plug into existing agents, with upcoming training options from multiple providers. We've included two agents we created under the Xeno account—Xenbot Truth Terminal and Xenobot bully—plus a 4chan-crypto model to demonstrate how longer-context models engage in conversation.
X-data: A dedicated index of highly curated, ready-to-use datasets and a toolkit for formatting and creating datasets from raw sources. We've added examples from our existing datasets covering social data and longer context conversations.
Check the repo to stay up-to-date with upcoming features.
With these components, our vision is to provide a drag-and-drop, mix-and-match system enabling anyone to create their own dataset and fine-tuned model, while offering more customization for advanced users.
I'm eager to get feedback from the community and builders and looking forward to promoting the adoption of the framework with the $XENO community!