Upload README.md with huggingface_hub

README.md

@@ -21,6 +21,7 @@ Each round — a **loop** — proposes new culture protocols, runs them at the b
 resulting cells. The measurements from each loop are published here as a separate file.
 
 - **Code and full reproduction pipeline:** <https://github.com/theislab/LabCompass>
+- **Contents:** per-loop measurements in `loops/`, trained models in `checkpoints/`
 - **Wet-lab experiments and measurements:** Göttgens Lab
 - **License:** CC-BY-4.0
 
@@ -112,6 +113,69 @@ path = hf_hub_download(
 adata = ad.read_h5ad(path)
 ```
 
+## Model checkpoints
+
+`checkpoints/` holds the trained models behind the paper's designs, one folder per loop:
+
+| Folder | Size | Forward model | Config group |
+| --- | --- | --- | --- |
+| `checkpoints/loop0/` | 1.7 GB | `likely-donkey-20` | `paths=loop0` |
+| `checkpoints/loop1/` | 1.9 GB | `eager-feather-1` | `paths=loop1` |
+| `checkpoints/loop2/` | 2.1 GB | `fresh-bee-21` | `paths=loop2_replicate` |
+| `checkpoints/loop2p5/` | 2.3 GB | `celestial-fire-40` | `paths=loop2p5_replicate` |
+| `checkpoints/loop3/` | 2.6 GB | `rich-sunset-44` | `paths=loop3_replicate` |
+| `checkpoints/loop4/` | 2.6 GB | `fast-gorge-13` | `paths=loop4_replicate` |
+
+**Loop *N*'s models are the ones that generated the designs executed in loop *N+1*.** So to reproduce
+the candidates that were run at the bench in loop 1, use `checkpoints/loop0/`.
+
+Each folder contains the four models the pipeline needs:
+
+- **`*_FlowMatching.pkl`** — the forward model: predicts the cell-state distribution a protocol induces.
+- **`*_TargetPredictionModel.pkl`** — the cell-type classifier, i.e. the phenotypic readout that the
+  inverse objective is defined against.
+- **`*_FlowMatchingWithScore.pkl`** — the generative prior over protocols.
+- **`*_FlowMap.pkl`** — a distilled few-step version of that prior.
+
+File names are the Weights & Biases run names, unchanged from training. `checkpoints/manifest.json`
+records, for every file, which `paths` config group refers to it, its original size, and a sha256.
+
+### Three things to know before using them
+
+**They are inference-only.** The training and validation data that the original checkpoints carried
+inside them has been removed — that is why a 113 GB file is 1 GB here. Everything inference touches
+is intact (network weights bitwise unchanged, normalisation parameters, cell-type labels, the
+condition key), and every checkpoint was verified tensor-by-tensor against its original and exercised
+end-to-end through the pipeline. But you **cannot retrain a prior from these**: the scripts that do
+so read the forward model's embedded training set, which is gone. Retrain the forward model from the
+`loops/` data instead.
+
+**They are CPU-resident.** The tensors load on any machine, with or without a GPU; move the model to
+your device as you would any PyTorch module. The original checkpoints held CUDA tensors and could
+only be loaded on a GPU node.
+
+**Loops 3 and 4 use the expanded design space.** They were trained after M-CSF and the lymphoid
+cocktail were added, so they expect the wider protocol vector and will fail with a shape mismatch if
+you load them with the earlier annotation. Use `annotation=bloodplus_loop3` for those two; the
+earlier loops use the default.
+
+### Loading
+
+```python
+from huggingface_hub import hf_hub_download
+from labcompass.models import FlowMatching
+
+path = hf_hub_download(
+    repo_id="theislab/LabCompass",
+    filename="checkpoints/loop3/rich-sunset-44_FlowMatching.pkl",
+    repo_type="dataset",
+)
+model = FlowMatching.load(path)
+```
+
+In the reproduction repository these are wired up through the `paths` config group, so pointing a run
+at a downloaded loop is a matter of overriding the four checkpoint paths.
+
 ## Citation
 
 <!-- TODO: replace with the published reference before release. -->