Update readme with download instructions

README.md

@@ -129,4 +129,61 @@ The dataloader uses this JSONL as the centralized index to locate and load all r
 4. Optionally load `.pt` behavioral feature tensors.
 5. Construct the multimodal sample for training or evaluation.
 
-This unified indexing structure enables heterogeneous behavioral datasets to be standardized under a single multimodal evaluation framework.
+This unified indexing structure enables heterogeneous behavioral datasets to be standardized under a single multimodal evaluation framework.
+
+---
+
+# Instructions to Download the Full Benchmark (JSONLs + Raw Media)
+
+The dataset consists of:
+
+- JSONL split files at the repository root
+- Multipart tar archive under `parts/`
+
+Files:
+
+```
+parts/human_behaviour_data.tar.part-000
+...
+parts/human_behaviour_data.tar.part-009
+```
+
+You must download **both** the JSONLs and the archive parts.
+
+---
+
+# Recommended Method — Using huggingface-cli
+
+If you have sufficient disk space, this is the simplest and safest approach:
+
+```bash
+pip install -U "huggingface_hub[cli]"
+huggingface-cli login
+
+huggingface-cli download keentomato/human_behavior_atlas \
+  --repo-type dataset \
+  --local-dir hba_download \
+  --local-dir-use-symlinks False
+```
+
+Then merge and extract:
+
+```bash
+cd hba_download
+cat parts/human_behaviour_data.tar.part-* > human_behaviour_data.tar
+tar -xf human_behaviour_data.tar
+```
+
+After extraction, your directory structure will look like:
+
+```
+hba_download/
+├── final_v8_train_cleaned_2.jsonl
+├── final_v8_val_cleaned.jsonl
+├── final_v8_test_cleaned.jsonl
+├── pose/
+├── opensmile/
+├── cremad_dataset_audio/
+├── ...
+```
+