Upload folder using huggingface_hub

DATASET_CARD.md

@@ -1,7 +1,6 @@
 ---
 license: mit
 task_categories:
-  - text2text-generation
   - question-answering
 language:
   - en
@@ -33,7 +32,7 @@ configs:
 
 Invoice processing is the use case every enterprise AI pitch deck opens with. The numbers are either right or wrong, and the distance between right and wrong can be measured to the cent. This dataset exists because we ran the experiment and discovered that the gap between "this looks easy" and "this actually works" is wider than the industry would like to admit.
 
-Five open-weight models. Four architectures. The best one scored 83%. The largest one scored 77%. The reasoning models performed worse than the plain models at every size. The full write-up is at [jngb.io/notes/06-too-dangerous-to-release](https://jngb.io/notes/06-too-dangerous-to-release).
+Five open-weight models. Four architectures. The best one scored 83%. The largest one scored 77%. The reasoning models performed worse than the plain models at every size. The full write-up is at [jngb.online/notes/06-too-dangerous-to-release](https://www.jngb.online/notes/06-too-dangerous-to-release).
 
 ## What's in the Box
 
@@ -186,7 +185,7 @@ If you use this dataset, please cite:
   title={InvoiceBenchmark: A Controlled Corpus for Measuring LLM Invoice Processing Accuracy},
   author={Neugebauer, Jakob},
   year={2026},
-  url={https://jngb.io/notes/06-too-dangerous-to-release},
+  url={https://www.jngb.online/notes/06-too-dangerous-to-release},
   note={200 synthetic invoices varying across five controlled dimensions with cent-perfect ground truth}
 }
 ```

README.md

@@ -1,7 +1,6 @@
 ---
 license: mit
 task_categories:
-  - text2text-generation
   - question-answering
 language:
   - en
@@ -33,7 +32,7 @@ configs:
 
 Invoice processing is the use case every enterprise AI pitch deck opens with. The numbers are either right or wrong, and the distance between right and wrong can be measured to the cent. This dataset exists because we ran the experiment and discovered that the gap between "this looks easy" and "this actually works" is wider than the industry would like to admit.
 
-Five open-weight models. Four architectures. The best one scored 83%. The largest one scored 77%. The reasoning models performed worse than the plain models at every size. The full write-up is at [jngb.io/notes/06-too-dangerous-to-release](https://jngb.io/notes/06-too-dangerous-to-release).
+Five open-weight models. Four architectures. The best one scored 83%. The largest one scored 77%. The reasoning models performed worse than the plain models at every size. The full write-up is at [jngb.online/notes/06-too-dangerous-to-release](https://www.jngb.online/notes/06-too-dangerous-to-release).
 
 ## What's in the Box
 
@@ -186,7 +185,7 @@ If you use this dataset, please cite:
   title={InvoiceBenchmark: A Controlled Corpus for Measuring LLM Invoice Processing Accuracy},
   author={Neugebauer, Jakob},
   year={2026},
-  url={https://jngb.io/notes/06-too-dangerous-to-release},
+  url={https://www.jngb.online/notes/06-too-dangerous-to-release},
   note={200 synthetic invoices varying across five controlled dimensions with cent-perfect ground truth}
 }
 ```