Add link to paper, task category

README.md

@@ -1,3 +1,8 @@
+---
+task_categories:
+- text-generation
+---
+
 # StreamProfileBench
 
 Streaming user-interest profiling benchmark for Chinese social media platforms.
@@ -5,6 +10,10 @@ Each task asks an LLM to (a) maintain a rolling persona summary from a stream
 of a user's posts, and (b) predict which tags from a curated candidate pool the
 user will engage with in the next time window.
 
+This repository contains the data presented in [StreamProfileBench: A Benchmark for Fine-Grained User Profile Inference in Real-World Streaming Scenarios](https://huggingface.co/papers/2605.25758).
+
+Code: https://github.com/WaterWang-001/StreamProfileBench
+
 The benchmark evaluates two coupled abilities:
 
 - **Plasticity** — picking up *newly emerging* interests (`Recall_Novelty`).
@@ -120,7 +129,9 @@ One JSON object per line in each `data/<platform>_inference_tasks.jsonl`:
       "total_steps": 4,
       "date_input":  "2025-06-12",
       "date_target": "2025-06-17",
-      "posts_text":  "[1] Content: ...\nTags: ...\n[2] ...",
+      "posts_text":  "[1] Content: ...
+Tags: ...
+[2] ...",
       "candidate_pool": ["tag1", "tag2", ...],
       "ground_truth": {
         "all_tags":  ["..."],
@@ -199,4 +210,4 @@ If you find any residual PII please open an issue.
 
 ## License
 
-Apache License 2.0. See [LICENSE](LICENSE).
+Apache License 2.0. See [LICENSE](LICENSE).