Update README.md

README.md

@@ -4,4 +4,27 @@ task_categories:
 - question-answering
 language:
 - en
----
+---
+
+This repo contains the raw data for the paper [**From Atomic to Composite: Reinforcement Learning Enables Generalization in Complementary Reasoning**](https://arxiv.org/pdf/2512.01970) with human biographies from a synthetic knowledge graph. 
+
+The code to generate these data is available at [https://github.com/sitaocheng/from_atomic_to_composite](https://github.com/sitaocheng/from_atomic_to_composite).
+
+The data can be directly adapted to frameworks like [LLamafactory](https://github.com/hiyouga/LLaMA-Factory) or [VeRL](https://github.com/volcengine/verl). 
+
+We opensource the training and testing data for parametric, contextual and complementary reasoning, respectively. 
+
+The training data is in the format of SFT full fine-tuning for LLamafactory. It contains the paramtric knowledge and QA-pairs for each reasoning type. For reinforcement learning (RL), we only adopt QA-pairs. To use VeRL, please adopt the data preprocess script to prepare the data.
+
+Example:
+```
+  "question": "Can you tell me the phone number of the boss of Justin Martinez's roommate?",
+  "answer": "6548113713",
+  "answer_cot": "Justin Martinez shared a room with Carl Jarvis. Frederick Juarez is the boss of Carl Jarvis. Frederick Juarez's phone number is 6548113713. So, the answer is: 6548113713",
+  "instruction": "Answer the following question.",
+  "gen_type": "composition"
+```
+
+The testing data is splited based on three level of generalization difficulty (i.e., I.I.D., composition, zero-shot generalization). The training data is sufficient to achieve excellent I.I.D. performance.
+
+