From e1bf8f9de984cd8ed992bb947efff07180d6ae24 Mon Sep 17 00:00:00 2001
From: Joya Chen <joyachen@u.nus.edu>
Date: Fri, 28 Jun 2024 17:56:27 +0800
Subject: [PATCH] Create README.md

---
 data/livechat/README.md | 35 +++++++++++++++++++++++++++++++++++
 1 file changed, 35 insertions(+)
 create mode 100644 data/livechat/README.md

diff --git a/data/livechat/README.md b/data/livechat/README.md
new file mode 100644
index 0000000..a7de91a
--- /dev/null
+++ b/data/livechat/README.md
@@ -0,0 +1,35 @@
+## Distributed Streaming Dialogue Data Generation
+
+This part describes how to generate streaming dialogue data on Ego4D GoalStep dataset.
+
+### Download Ego4D GoalStep Annotation JSON
+
+Try to download Ego4D annotations. Refer to [Ego4D](https://ego4d-data.org/docs/start-here/) for details.
+
+After that, you can use symbolic link to ensure you have ego4d annotations as the following:
+
+```
+datasets/ego4d/v2/annotations/
+├── ...
+├── goalstep_train.json
+├── goalstep_val.json
+├── ...
+```
+
+### Run Streaming Data Generation Script
+
+```
+python -m data.ego4d.livechat.ego4d_goalstep_livechat_generation --num_gpus 8 --anno_root 
+```
+
+- Please run the script in ```videollm-online/``` root folder.
+
+- If you are on a cluster, you can set ```--num_nodes ... --slurm_partition ...``` to use them. The more nodes and GPUs, the faster preprocessing.
+
+- The results will be saved into ```datasets/ego4d/v2/annotations/livechat/```. 
+
+### Filtering out Data expose ground-truth
+
+We find sometimes the generated dialogue will expose the timestamp information in ground-truth annotation. The pattern is "second", "..s" appeared in generated assistant responses. Furthermore, we remove dialogue less than 1 minute and larger than 60 minutes.
+
+See python [filter.py](filter.py) for details.