Update README.md

README.md

@@ -28,8 +28,30 @@ This simple donut isn't just a snack on a stick—it's proof that robots can lea
 
 Video (robot trained using this dataset, ACT policy for 40k steps): [link](https://huggingface.co/datasets/LeRobot-worldwide-hackathon/submissions/discussions/11)
 
+Train from scratch:
 
-Recording config:
+```
+python lerobot/scripts/train.py \
+    --dataset.repo_id=siakovlev/forloop_donut_on_stick_050 \
+    --policy.type=act \
+    --output_dir=outputs/train/forloop_donut_on_stick_policy_050 \
+    --job_name=forloop_job \
+    --policy.device=cuda
+```
+
+Train (fine-tune existing policy):
+
+```
+python lerobot/scripts/train.py \
+    --dataset.repo_id=siakovlev/forloop_donut_on_stick_020_val \
+    --policy.path=siakovlev/forloop_donut_on_stick_policy_050 \
+    --output_dir=outputs/train/forloop_donut_on_stick_policy_020 \
+    --job_name=forloop_job \
+    --policy.device=cuda \
+    --save_freq=4000 \
+```
+
+Recording config (for data recording and evaluation):
 
 ```
 python -m lerobot.record \