Saarthi (सारथी) — laya fine-tuned for browser use

Saarthi (Hindi for charioteer — steers and guides toward the goal) fine-tunes convaiinnovations/laya to pick the next browser operation (click / type_text / select) and target element in one non-autoregressive forward pass. Open replacement for the proprietary decision model in browser-use/jev-ultrafast.

Training data

osunlp/Mind2Web — real action traces across 137 sites. Each step -> operation-choice + target-choice (pos + sampled negatives).

Usage

import laya
agent = laya.load("Himvans/Saarthi")
agent.predict(state, questions)  # state has task+elements; questions are typed choices

type_text values need a small LLM (laya is a classifier), same as jev-ultrafast.

License

Apache-2.0 (inherits base).