Add link to paper and project page

README.md

@@ -1,21 +1,21 @@
 ---
-license:
-  - odc-by
-  - cc-by-nc-4.0
 language:
-  - en
+- en
+license:
+- odc-by
+- cc-by-nc-4.0
+size_categories:
+- 100K<n<1M
 task_categories:
-  - text-generation
-tags:
-  - safety
-  - data-attribution
-  - data-filtering
-  - gradient-embeddings
-  - tulu
-  - sft
+- text-generation
 pretty_name: SafeTulu
-size_categories:
-  - 100K<n<1M
+tags:
+- safety
+- data-attribution
+- data-filtering
+- gradient-embeddings
+- tulu
+- sft
 configs:
 - config_name: safe_tulu
   default: true
@@ -40,6 +40,9 @@ configs:
 
 > ⚠️ **Content warning:** `saferlhf_with_metadata` contains harmful requests and harmful responses taken from PKU-SafeRLHF 
 
+**Paper:** [Follow the Gradient: Tracing Safety Degradation Back to Training Data](https://huggingface.co/papers/2610.37481)  
+**Project page:** [DatologyAI blog](https://www.datologyai.com/blog/follow-the-gradient)
+
 SafeTulu is a version of the [Tulu 3 SFT mixture](https://huggingface.co/datasets/allenai/tulu-3-sft-mixture) with the examples most likely to harm a model's safety removed. Each Tulu 3 example is scored by how close its training gradient is to the gradients of harmful [PKU-SafeRLHF](https://huggingface.co/datasets/PKU-Alignment/PKU-SafeRLHF) examples, and the highest-scoring examples are removed.
 
 Along with the filtered dataset, we release the gradient embedding of every Tulu 3 and SafeRLHF example and the score of every Tulu 3 example. This lets you build your own version of SafeTulu. Some customization examples: keeping and filtering only non-English data, removing only 10% of the data, etc.