Aptheos/lfm-browser-control-evidence

Browser control evidence

Raw run reports behind browser-control: every evaluation episode, with observations, raw model outputs, tokens where recorded, native rewards, source hashes, the SFT expert corpus, and the vision screenshots.

# From a clone of the code repository:
uvx --from huggingface_hub hf download Aptheos/lfm-browser-control-evidence --repo-type dataset --local-dir evidence
python scripts/verify_reported_numbers.py --evidence evidence

Folder	Contents
lfm3-book-flight/	Base gate and its failed first attempt, a base run matched to the SFT development cohort (base-development-matched/), development diagnostic, all SFT stages and the rejected pass, reserved evaluation, GRPO runs and gates, inputs
lfm3-click-test/	Frozen LFM3 click-test
lfm2-click-test/	Three GRPO seed reports, and the decoder re-score
release-check/	2–3 November rerun of the published commands and adapters from a clean copy
vision/frozen-matrix/, vision/variation/	LFM3-VL-2B base matrix and variation reports, with screenshots
vision/click-color-fit/	One-click fit: expert labels, SFT and GRPO training records, development evaluation
vision/click-color-reserved/	Locked protocol and reserved evaluation of base, SFT and GRPO on click-color
vision/click-sequence-baseline/	Frozen multi-step baseline on click-button-sequence (the SFT variation step stopped on a fixed bug and is kept as recorded)
vision/click-sequence-fit/	Sequence expert labels, SFT training record, development evaluation and variation
vision/click-sequence-reserved/	Locked protocol and reserved evaluation of base and the sequence adapter
vision/click-sequence-grpo/	Simplified GRPO from the sequence adapter: contract, 120 training groups with rewards and gradients, development evaluation and variation
vision/click-sequence-reserved-b/	Second locked protocol and reserved evaluation, seeds 22000–22099: sequence SFT against SFT then GRPO

These are development records, not a benchmark dataset. All tasks, flights and airports come from MiniWoB's simulated pages (MIT, Copyright 2016 OpenAI). Model outputs in the reports came from Liquid AI's LFM models, and the reports' reasoning text is the models' own. Commit hashes refer to the development history; most LFM3 reports also record the SHA-256 of its source files.