Add paper link to dataset card

README.md

@@ -1,18 +1,20 @@
 ---
 license: mit
 task_categories:
-  - visual-question-answering
-tags:
-  - medical
-  - ct
-  - 3d
-  - radiology
-  - vqa
+- visual-question-answering
 pretty_name: MeasureVQA
+tags:
+- medical
+- ct
+- 3d
+- radiology
+- vqa
 ---
 
 # MeasureVQA
 
+Paper: [When Do Cheap Probes Predict Expensive Training? Probing 3D-CT Encoders for Text Generation](https://huggingface.co/papers/2607.22771)
+
 An **image-grounded** measurement VQA benchmark for 3D CT, released with
 **CheapCT** ([code](https://github.com/renjie-liang/CheapCT)).
 
@@ -57,7 +59,10 @@ Each file is a list of VQA records in LLaVA-style conversation format:
   "answer_letters":["C"],                         // gold letter(s)
   "flag":          "",                            // e.g. MARGINAL for the spatial-position canary
   "conversations": [
-    {"from": "human", "value": "<image>\nHow would you characterize the cardiothoracic ...?\n(A) High (B) Moderate (C) Low\nAnswer with the letter only."},
+    {"from": "human", "value": "<image>
+How would you characterize the cardiothoracic ...?
+(A) High (B) Moderate (C) Low
+Answer with the letter only."},
     {"from": "gpt",   "value": "C"}
   ]
 }
@@ -80,4 +85,4 @@ here. Obtain the raw CT volumes from their sources under their own licenses:
 ```python
 import json
 records = json.load(open("ctrate/valid_size.json"))
-```
+```