update README for per-model layout

README.md

@@ -5,13 +5,16 @@ tags: [jax, flax, cifar10, mnist, slice-sampling]
 
 # slice-sampling checkpoints
 
-Auxiliary checkpoints for the 'outsourced' experiments of Generalised Latent Slice Sampling progect.
+Auxiliary checkpoints for the `slice_sampling` research codebase (`outsourced/`
+targets that build MCMC benchmarks over pretrained generator + classifier
+stacks). Each subdirectory is self-contained: the checkpoint, the model source
+file, and a small `load.py` that reconstructs the model from the pickled meta.
 
-| file | arch | dataset |
-|------|------|---------|
-| `cifar10_classifier.pkl` | `Cifar10CNN` (~1.15M params, GroupNorm, 32x32 native) | CIFAR10 |
-| `mnist_classifier.pkl`   | `Classifier` (2 conv stages + FC, 32x32) | MNIST |
-| `mnist_vae.pkl`          | ResNet-style VAE, `latent_dim=16` | MNIST |
+| directory | model | dataset | notes |
+|-----------|-------|---------|-------|
+| `cifar10_classifier/` | `Cifar10CNN` (~1.15M params, GroupNorm) | CIFAR10 | 32x32 native, no upsample |
+| `mnist_classifier/` | `Classifier` (2 conv stages + FC) | MNIST | 32x32 (padded) |
+| `mnist_vae/` | ResNet-style VAE, `latent_dim=16` | MNIST | encoder/decoder both included |
 
 Fetched automatically by `outsourced.common.load_checkpoint` when the local
 file is missing.