Update dataset card with paper/code links and task categories

README.md

@@ -1,29 +1,34 @@
 ---
-pretty_name: PrimeKG-CL
-license: cc-by-4.0
-task_categories:
-- other
 language:
 - en
+license: cc-by-4.0
+size_categories:
+- 10M<n<100M
+task_categories:
+- graph-ml
+- question-answering
+pretty_name: PrimeKG-CL
 tags:
 - biomedical
 - knowledge-graph
 - continual-learning
 - link-prediction
-size_categories:
-- 10M<n<100M
 ---
 
 # PrimeKG-CL
 
-PrimeKG-CL is a continual graph learning benchmark built on an evolving biomedical knowledge graph. It contains two temporal snapshots (`t0`, `t1`), temporal diffs, 10 continual-learning tasks, and multimodal node features for evaluating retention/forgetting under distribution shift.
+PrimeKG-CL is a continual graph learning benchmark built on an evolving biomedical knowledge graph. It was introduced in the paper [PrimeKG-CL: A Continual Graph Learning Benchmark on Evolving Biomedical Knowledge Graphs](https://huggingface.co/papers/2605.10529).
+
+- **Code:** [https://github.com/yradwan147/primekg-cl-neurips2026](https://github.com/yradwan147/primekg-cl-neurips2026)
+
+PrimeKG-CL contains two temporal snapshots (`t0`, `t1`), temporal diffs, 10 continual-learning tasks, and multimodal node features for evaluating retention/forgetting under distribution shift.
 
 This dataset repository stores the benchmark under the `benchmark/` directory.
 
 ## Dataset Summary
 
 - **Domain:** Biomedical knowledge graphs
-- **Primary use:** Continual knowledge graph completion / link prediction
+- **Primary use:** Continual knowledge graph completion / link prediction / KGQA
 - **Snapshots:**
   - `benchmark/snapshots/kg_t0.csv` (June 2021 PrimeKG release): 8,100,498 triples, 129,375 nodes, 30 relations
   - `benchmark/snapshots/kg_t1.csv` (July 2023 reconstruction): 13,001,666 triples, 134,211 nodes, 25 relations
@@ -51,11 +56,17 @@ This dataset repository stores the benchmark under the `benchmark/` directory.
 - Task splits use 70/10/20 train/valid/test.
 - Feature tensors are stored as `.pt` files.
 
-## Suggested Usage
+## Sample Usage
 
-1. Use `task_0_base` as the base pretraining stage on `t0`.
-2. Train/evaluate sequentially on tasks `task_1_*` through `task_5_*`.
-3. Report both aggregate and stratified metrics (persistent vs added vs removed), using `benchmark/test_stratification.json`.
+After setting up the environment using the `environment.yml` from the GitHub repository, you can run a baseline (e.g., EWC on DistMult) using the following command:
+
+```bash
+python scripts/run_baselines.py \
+    --method ewc --decoder DistMult \
+    --seeds 42 123 456 789 1024 \
+    --data_root data/benchmark \
+    --output results/
+```
 
 ## Limitations and Considerations
 
@@ -81,4 +92,4 @@ If you use this dataset, please cite:
   booktitle={NeurIPS Datasets and Benchmarks Track},
   year={2026}
 }
-```
+```