transformers remote code: AutoModel.from_pretrained('cklxx/laya-browser', trust_remote_code=True).decide(page, goal)

Root config.json (lets the Hub count downloads) + library_name: laya

README: hf download (huggingface-cli is deprecated)

laya_browser.py: no ModernBERT auto-compile (works without a Triton toolchain); README uses `hf download`

Model at the repo root + laya_browser.py: load with LayaBrowser.from_pretrained('cklxx/laya-browser') or serve a TypeSafe endpoint

v19s: WebChain real-site trajectories, format v5, webgym x7 + DAgger, harness fixes; replaces v17s

Single model: v17s (suite A 85 %, suite B 100 %, webgym 47 %); replaces v15s (kept in history)

Single model: v15s (clean retrain, suite B 100 %); remove v10, v10s, v11s, v14s (kept in history)

v14s checkpoint, held-out suite B (100 %), harness fixes, webgym + sub-goal data pipeline, DAgger leak correction

kernels: sliding window = config.sliding_window (64, not 65); fp32 residual stream; parity_fast.py (mirrors laya PR #25 commit 2e54f87)

README: final suite numbers, teacher comparison, v11s

results: corrected suite x3 for v10/v10s/v11s, teacher comparison

code: rollouts, teacher eval, corrected suite checks

results: v11s suite

v11s: 322M + scripted scroll/search/select trajectories; 58% on 16 real tasks x3 at 19 ms/step

README: demo gif

code: demo recorder

assets: demo video + gif

code: uv project (pyproject + uv.lock), verify.py, sample_request.json, fast-path import fix

laya-browser v10 / v10s: laya fine-tuned as a browser-agent decision head + code + results

initial commit