Extend docs

docs/beware.md

@@ -35,3 +35,21 @@ you should do two things:
 1. Set `torch.use_deterministic_algorithms(True)` to make PyTorch operations deterministic.
 
 2. Set `model.eval()` to disable drop-out.
+
+## Loading a Checkpoint Onto an Extended Model
+
+If you changed the model and added or removed parameters, you need to set `strict=False` when
+loading a checkpoint `Aurora.load_checkpoint(..., strict=False)`.
+Importantly, enabling or disabling LoRA for a model that was trained respectively without or
+with LoRA changes the parameters!
+
+## Extending the Model with New Surface-Level Variables
+
+Whereas we have attempted to design a robust and flexible model,
+inevitably some unfortunate design choices slipped through.
+
+A notable unfortunate design choice is that extending the model with a new surface-level
+variable breaks compatibility with existing checkpoints.
+It is possible to hack around this in a relatively simple way.
+We are working on a more principled fix.
+Please open an issue if this is a problem for you.

docs/finetuning.md

@@ -10,4 +10,21 @@ model = Aurora(use_lora=False)  # Model is not fine-tuned.
 model.load_checkpoint("wbruinsma/aurora", "aurora-0.25-pretrained.ckpt")
 ```
 
-More specific instructions coming soon.
+You are also free to extend the model for your particular use case.
+In that case, it might be that you add or remove parameters.
+Then `Aurora.load_checkpoint` will error,
+because the existing checkpoint now mismatches with the model's parameters.
+Simply set `Aurora.load_checkpoint(..., strict=False)`:
+
+```python
+from aurora import Aurora
+
+
+model = Aurora(...)
+
+... # Modify `model`
+
+model.load_checkpoint("wbruinsma/aurora", "aurora-0.25-pretrained.ckpt", strict=False)
+```
+
+More instructions coming soon!