Clarify CueBack data variants

README.md

@@ -43,10 +43,17 @@ embodied vision-language models, built on
 
 - `data/train/`: 3,400 training tasks.
 - `data/test/`: 1,000 human-validated evaluation tasks.
-- `cueback/sft_action.jsonl`: executable action trajectories without reasoning.
-- `cueback/sft_both.jsonl`: trajectories with directly generated reasoning.
-- `cueback/cueback.jsonl`: trajectories with CueBack-constructed reasoning.
-- `cueback/images/`: referenced first-person observations.
+
+The CueBack data contains 3,158 aligned trajectories in three variants. They
+share the same first-person observations, tool calls, and tool feedback, but
+differ in the supervision text preceding each action:
+
+- `cueback/sft_action.jsonl`: executable actions without reasoning.
+- `cueback/sft_both.jsonl`: the same actions with rationales generated directly
+  from the trajectory prefix, current observation, and action.
+- `cueback/cueback.jsonl`: the same actions with clue-based reasoning
+  constructed by CueBack.
+- `cueback/images/`: the referenced first-person observations.
 
 The benchmark format, setup instructions, and evaluation code are documented
 in the [HomeSafeBench code repository](https://github.com/BITHLP/HomeSafeBench).