diff --git a/.readthedocs.yaml b/.readthedocs.yaml
index 31941bc..c1b81e4 100644
--- a/.readthedocs.yaml
+++ b/.readthedocs.yaml
@@ -29,4 +29,4 @@ sphinx:
 # See https://docs.readthedocs.io/en/stable/guides/reproducible-builds.html
 python:
     install:
-    - requirements: ci_requirements.txt
+    - requirements: doc/requirements.txt
diff --git a/README.md b/README.md
index 87cfc21..4f992df 100644
--- a/README.md
+++ b/README.md
@@ -26,7 +26,7 @@ pip install git+https://github.com/tangentlabs/django-oscar-paypal.git@issue/34/
 ```
 
 ## Minimal Example
-[examples/barebone_mnist.py](example/barebone_mnist.py) features a minimal and barebone example on how to distributely train MNIST.
+See [examples/barebone_mnist.py](https://github.com/sehoffmann/dmlcloud/blob/develop/examples/barebone_mnist.py) for a minimal and barebone example on how to distributely train MNIST.
 To run it on a single node with 4 GPUs, use
 ```
 dmlrun -n 4 examples/barebone_mnist.py
diff --git a/dmlcloud/run.py b/dmlcloud/run.py
index 0f56aec..ac119b2 100644
--- a/dmlcloud/run.py
+++ b/dmlcloud/run.py
@@ -23,15 +23,20 @@
 import argparse
 import os
 
+
 def main():
-    description = ('dmlrun is a thin wrapper around torch.distributed.launch that provides a more user-friendly interface.\n\n'
-                    'While torchrun is a powerful tool, it can be a bit clunky to use for testing and debugging. dmlrun aims to make it easier to launch distributed training jobs on a single node.'
-                    'For serious mulit-node training, we recommend using srun or torchrun directly.')
-    epilog = ('Example:\n'
-                '    dmlrun --gpus 3,7 train.py\n'
-                '    dmlrun --num-gpus 2 train.py --batch-size 64')
-    parser = argparse.ArgumentParser(prog='dmlrun', description=description, epilog=epilog, formatter_class=argparse.RawDescriptionHelpFormatter)
-    parser.add_argument('--gpus', '-g', help='Comma-seperated list of GPU IDs to use for training. Overrides CUDA_VISIBLE_DEVICES.')
+    description = (
+        'dmlrun is a thin wrapper around torch.distributed.launch that provides a more user-friendly interface.\n\n'
+        'While torchrun is a powerful tool, it can be a bit clunky to use for testing and debugging. dmlrun aims to make it easier to launch distributed training jobs on a single node.'
+        'For serious mulit-node training, we recommend using srun or torchrun directly.'
+    )
+    epilog = 'Example:\n' '    dmlrun --gpus 3,7 train.py\n' '    dmlrun --num-gpus 2 train.py --batch-size 64'
+    parser = argparse.ArgumentParser(
+        prog='dmlrun', description=description, epilog=epilog, formatter_class=argparse.RawDescriptionHelpFormatter
+    )
+    parser.add_argument(
+        '--gpus', '-g', help='Comma-seperated list of GPU IDs to use for training. Overrides CUDA_VISIBLE_DEVICES.'
+    )
     parser.add_argument('--nprocs', '-n', type=int, help='Number of GPUs to use for training.')
     parser.add_argument('script', type=str, help='Path to the script to run.')
     parser.add_argument('args', nargs=argparse.REMAINDER, help='Arguments to pass to the script.')
@@ -41,7 +46,6 @@ def main():
     if args.gpus and args.num_gpus:
         raise ValueError('Only one of --gpus or --num-gpus can be specified.')
 
-    
     if args.gpus:
         ids = args.gpus.split(',')
         if not all(id.isdigit() for id in ids):
@@ -67,4 +71,4 @@ def main():
 
 
 if __name__ == '__main__':
-    main()
\ No newline at end of file
+    main()
diff --git a/doc/requirements.txt b/doc/requirements.txt
new file mode 100644
index 0000000..e7e6280
--- /dev/null
+++ b/doc/requirements.txt
@@ -0,0 +1,3 @@
+-r ../ci_requirements.txt
+-e ./
+-r ../requirements.txt