Clean retrain checkpoint (used by the shortlist test); strict loading; card fixes

README.md

@@ -16,10 +16,10 @@ Paper: *A 149M Text Encoder as the First Tier for Phone Agents: Pre-Registered R
 - **What did not work is documented too.** About 40 registered experiments, negative ones included, with timestamps
   (`paper/PREREGISTRATION.md`).
 
-All Claude API calls behind the AndroidControl results cost **$89.73** in total (Batch prices); each training run takes about
+All Claude API calls behind the AndroidControl results cost **$86.75** in total (Batch prices); each training run takes about
 **6.4 hours on one NVIDIA T4**.
 
-![Cost and accuracy on 1,000 held-out steps; routing curves](fig_router.png)
+![Cost and accuracy on 1,000 held-out steps; routing curves](https://huggingface.co/coolnow/milo-pointer/resolve/main/fig_router.png)
 
 Demo: the model drives a Pixel 6a through a five-step Clock task ("set a 5 minute timer"), every step answered locally
 ([video](https://huggingface.co/coolnow/milo-pointer/resolve/main/milo_demo_clean.mp4)).

clean_r3/pointer_clean_r3.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5a5a2d7c9801e7db4d2ff24321af24eb1a88a5dd414dca9f927346a6ef2f442b
+size 596158793

milo_pointer.py

@@ -96,7 +96,7 @@ class MiloPointer:
         cfg = AutoConfig.from_pretrained(p if (p / "config.json").exists() else BASE)
         net = _Net(cfg)
         net.load_state_dict(
-            {k: v.float() for k, v in load_file(str(p / "model.safetensors")).items()}, strict=False
+            {k: v.float() for k, v in load_file(str(p / "model.safetensors")).items()}, strict=True
         )
         net = net.to(device).eval()
         if device != "cpu":