Upload README.md with huggingface_hub

README.md

@@ -1,18 +1,23 @@
 ---
 license: apache-2.0
 library_name: laya
+base_model: convaiinnovations/laya
 tags: [laya, system-one, android, phone-agent, decision-model]
 ---
 
 # laya-ac-phone-v1
 
-Laya-english (ModernBERT-large, 421M) fully finetuned with RLCD on AndroidControl
-phone-control steps converted to System One decisions (`operation` choice over
-TAP/TYPE/SCROLL/BACK/WAIT/DONE + `done` noul; state = goal + step history).
+Full RLCD finetune of [convaiinnovations/laya](https://huggingface.co/convaiinnovations/laya)
+(ModernBERT-large, 421M) for Android phone control.
 
-- Train: 18,354 rows, 4 epochs on Kaggle T4 (`eros438/phone-decisions-ac`)
-- Held-out operation accuracy (984 unseen goals): base 0.261 → **0.665**
-- In-sample calibration: acc 0.85, ECE 0.037
+**Training data:** [eros438/phone-decisions-ac](https://www.kaggle.com/datasets/eros438/phone-decisions-ac) —
+18,354 decision rows converted from [xwm/AndroidControl](https://huggingface.co/datasets/xwm/AndroidControl)
+(a parquet mirror of Google's AndroidControl: 15,283 human demos, 833 apps).
+Each step became an `operation` choice (TAP/TYPE/SCROLL/BACK/WAIT/DONE, state = goal + step
+history) plus a `done` noul. 4 epochs on a Kaggle T4 via `laya-train` (defaults).
+
+**Results (984 held-out goals):** operation accuracy base 0.261 → **0.665**;
+in-sample calibration acc 0.85, ECE 0.037.
 
 ```python
 import laya
@@ -23,4 +28,4 @@ agent.predict("Goal: ...", {"operation": {"type": "choice",
                  "BACK": "...", "WAIT": "...", "DONE": "..."}}})
 ```
 
-Serves via `laya-serve` (`LAYA_MODELS` pointing at this repo).
+Serves via `laya-serve` pointed at this repo.