WebLINX

WebLINX is a benchmark of real-world website navigation demonstrations collected as multi-turn dialogues between a human instructor and a human navigator. Per the paper, it comprises roughly 2,300 expert demonstrations spanning about 100K interaction steps across 150+ real websites, covering everyday tasks such as booking flights, shopping, and managing accounts. Each demonstration is decomposed into turn-level instances: given the dialogue utterances so far, the prior action history, a cleaned/pruned HTML DOM, and a list of candidate elements, a model must produce the reference navigator action in WebLINX's action DSL (e.g. click(uid=...), say(speaker=..., utterance=...), load(url=...), submit(uid=...), change(...), scroll()). The Hugging Face release exposes a text-based 'chat' config (24,418 train turn-level rows plus validation and multiple out-of-domain test splits) and a 'reranking' config for candidate-element selection. Models are evaluated per turn against the demonstrated reference action using intent-match, element-overlap, and text-similarity metrics, including out-of-distribution generalization to unseen websites, categories, geographies, and visual layouts.