Fine-tuning Liquid d1 decision models on screenshots: 51% to 96% accuracy in ~20 minutes
helloiamleonie · x · 2026-10-30
Chris Levy extended Leonie's text-only recipe for fine-tuning Liquid d1 decision models to image inputs. Task: given a website screenshot, a task and candidate page elements, pick the next element to interact with (web-agent decisions), trained on Multimodal-Mind2Web (4,574 examples, dataset on Hugging Face). One epoch on a single RTX PRO 6000 via Modal:
More from coding & agent
- Scanning 25,267 MCP servers found 73% vulnerable, 143.5K vulnerabilities — anacondainc · 2026-10-30
- Carmack: AI writing software from scratch isn't optimal — well-regularized building blocks matter — ID_AA_Carmack · 2026-10-30
- LangChain Core 1.7.2 ships notification metadata methods and merge_content fix — caolonghua · 2026-10-30
- Codex on a phone: managing multiple coding agents remotely, shown in new screenshots — Dimillian · 2026-10-30
- Arm outlines native Arm dev workflow with GitHub Copilot CLI and Arm64 CI — caolonghua · 2026-10-30
- evident-charts: open-source skill teaches coding agents to make honest charts and self-check — randal_olson · 2026-10-30