Add task category, paper and code links

README.md

@@ -6,9 +6,15 @@ configs:
   data_files: miniswe.jsonl
 - config_name: terminus-2
   data_files: terminus-2.jsonl
+task_categories:
+- text-generation
 ---
+
 # SkillGym SFT trajectories
 
+Paper: [SkillGym: Training Skill-Use Agents with Automatic Verifiable Environment Generation](https://huggingface.co/papers/2609.37539)  
+Code: [https://github.com/Reason-Wang/SkillGym](https://github.com/Reason-Wang/SkillGym)
+
 19,053 successful (reward 1.0) agent trajectories on skill-requiring tasks, one JSONL file per agent harness.
 
 | file | harness | trajectories |
@@ -30,4 +36,4 @@ configs:
 Each message has `role` and `content`. `system` messages carry `tools` (the tool definitions); `assistant`
 messages carry `think` (reasoning) and `tool_calls` (`name`, `arguments`, `id`); `tool` messages carry
 `tool_call_id`. terminus-2 trajectories have no tool calls: the assistant replies with a JSON object and the
-terminal output returns as a `user` message.
+terminal output returns as a `user` message.