🚀 Tensorplex Dojo (Subnet 52) in action! Meet DOJO-INTERFACE-CODER-7B: Qwen2.5-Coder-7B-Instruct, fine-tuned with Dojo datasets to craft stunning front-end UIs! ✨ Generates beautiful, interactive interfaces ✨ Trained on synthetic data with distributed human feedback ✨ Powered by (Subnet 52) on Bittensor 👇

Here's how it works: - Validators on Dojo network generate diverse UI outputs using advanced AI models. - Human evaluators (miners) rate these UIs based on aesthetics, interactivity, and alignment with the intended task. - Feedback is collected into specialized datasets (SFT and DPO) to further enhance training. 2/8

Initially, Qwen2.5-Coder-7B-Instruct struggled significantly with generating complete UI code. We significantly improved its capabilities by training it on our high-quality 25k-completion SFT dataset, turning it into a reliable, structured interface generator. 3/8

Adding a 12.5k-completion DPO dataset evaluated by human contributors significantly improved UI alignment with real user preferences. The DPO-trained model clearly outperformed the SFT version in human evaluations. 4/8

Surprisingly, human-driven training (DPO) also boosted performance on general coding benchmarks like HumanEval and MBPP, even though training focused solely on UI tasks. 5/8

Potential use-cases for DOJO-INTERFACE-CODER-7B: - Adaptive educational interfaces - Privacy-centric customizable journaling tools - Dynamic UIs for enhanced human-AI collaboration 6/8

Next steps for Dojo Network: - Expanding our human-feedback loop for richer data collection - Developing dynamic human-agent interfaces - Building strategic partnerships across various industries 7/8