update experiment readme

.gitignore

@@ -168,4 +168,5 @@ wandb/
 
 experiments/*
 !experiments/.gitkeep
-!experiments/demo*
+!experiments/demo*
+!experiments/README.md

experiments/README.md

@@ -0,0 +1,58 @@
+# DEMO
+
+## data format
+- save data use parquet format
+- the actual data should be saved as list of dict under `messages` column
+- see `https://huggingface.co/datasets/HuggingFaceH4/SystemChat?row=0` as example
+
+sample code for data process:
+```python
+from datasets import load_dataset, Dataset
+
+# prepare data and save to local
+my_data = [
+    {
+        "messages": [
+            {"role": "system", "content": "This is system msg1"}, 
+            {"role": "user", "content": "This is user msg1"},
+            {"role": "assistant", "content": "This is assistant msg1"}, 
+        ]
+    },
+     {
+        "messages": [
+            {"role": "system", "content": "This is system msg2"}, 
+            {"role": "user", "content": "This is user msg2"},
+            {"role": "assistant", "content": "This is assistant msg2"}, 
+        ]
+    },
+]
+
+Dataset.from_list(my_data).to_parquet("my_dataset/train/data.parquet")
+Dataset.from_list(my_data).to_parquet("my_dataset/test/data.parquet")
+
+# load
+ds = load_dataset("my_dataset")
+print(ds)
+"""
+DatasetDict({
+    train: Dataset({
+        features: ['messages'],
+        num_rows: 2
+    })
+    test: Dataset({
+        features: ['messages'],
+        num_rows: 2
+    })
+})
+"""
+```
+
+## ENV
+- update the singulairty container path in `demo_dgx2_launcher.sh`
+- update python virtual environment path in `demo_dgx2.sh`
+
+## config
+- accelerate configuration is at `recipes/accelerate_configs`
+    - recommend to use deepspeed config to launch experiment
+- llama3-8b experiment config is under `recipes/llama3-8b`
+- you can put config whereever you like