deltas: document the three curated case-study deltas (remove forthcoming placeholder)

deltas/README.md

@@ -1,10 +1,78 @@
-# Curated case-study deltas — forthcoming
+# Curated case-study MEMIT deltas
 
-This directory is a placeholder. Curated MEMIT edit deltas for the paper's
-case studies (Eiffel Tower Paris->Rome, and Taj Mahal / Pentium II if
-available) are planned for a follow-up drop and are not included yet.
+This directory holds dense per-layer weight deltas for the paper's three
+worked case-study edits, each applied on top of `google/gemma-2-2b` (base,
+gated, subject to the [Gemma Terms of Use](https://ai.google.dev/gemma/terms))
+with [MEMIT](https://rome.baulab.info/) via the Interp_Gemma editing
+primitives. Applying a delta to the base weights reconstructs the edited
+model `M_e = base google/gemma-2-2b weights + delta`.
 
-When added, each delta reconstructs an edited model M_e as:
-`base google/gemma-2-2b weights + delta -> M_e`.
+No model weights or edited checkpoints are included — only the delta, which
+is meaningless without the (separately downloaded, license-gated) base model.
 
-No model weights or edited checkpoints are included in this release.
+## Files
+
+| Case | Subject → target (relation) | CounterFact `case_id` | File | Size | SHA-256 |
+|---|---|---|---|---|---|
+| Eiffel Tower Paris→Rome | *The Eiffel Tower* is located in the city of **Rome** (was Paris) | 0 | `eiffel_paris_to_rome/deltas.npz` | 679,480,828 B (~648 MiB) | `476e5fb6a4f660faf5472fbaf0a2eefb2c4bb58bf0dfbb5e0ad50e3230f5507c` |
+| Pentium II Intel→Microsoft | *Pentium II* was developed by **Microsoft** (was Intel) | 6467 | `pentium_ii_intel_to_microsoft/deltas.npz` | 679,480,800 B (~648 MiB) | `380ab17ff6524e2273dc1f68019601baee4acd3b047f35bd1f0106adeadc44fa` |
+| Taj Mahal Islam→Christianity | *Taj Mahal* is follower of **Christianity** (was Islam) | 21243 | `taj_mahal_islam_to_christianity/deltas.npz` | 679,480,804 B (~648 MiB) | `adeec83a4af118b1c34c6e2216ed9839eed351dcc002dc038a48070b081e9ca9` |
+
+The Eiffel Tower case is the bundle backing `bundles/eiffel_paris_to_rome/`
+elsewhere in this dataset and the library's shipped example bundle
+(`circuit_diff/examples/bundles/eiffel/`). Pentium II and Taj Mahal are two
+of the 24 edits in the `cf-sweep-24` campaign (see `tables/cf-sweep-24/`);
+their `case_id`s match the `case_id` column in that campaign's
+`edits.csv` / `setdiff_rows.csv`.
+
+This is a curated 3-edit subset chosen for the paper's case-study figures,
+not the full ~100-edit delta set used across the sweep campaigns; the full
+set is not planned for release but exists and can be shared on request.
+
+## npz key structure
+
+Each `deltas.npz` is a plain NumPy archive (`np.load(path)`) with:
+
+- `meta` — a 0-d array wrapping a single JSON string. Fields: `edit_method`
+  (`"memit"`), `request` (the MEMIT edit request: `case_id`, `prompt`
+  template, `subject`, `target_new.str`), `band` (the list of 0-indexed
+  transformer layers MEMIT was allowed to edit), `base_model`
+  (`"google/gemma-2-2b"`).
+- `model.layers.{L}.mlp.down_proj.weight` for each `L` in `band` (here,
+  `L` in `3..10`, one array per layer) — the dense weight delta
+  `w_edited − w_base` for that layer's MLP down-projection, `float32`,
+  shape `(2304, 9216)` (gemma-2-2b's hidden size × MLP intermediate size).
+  MEMIT only ever touches this one weight matrix per edited layer, so these
+  eight arrays fully specify the edit.
+
+## Reconstruction recipe
+
+```python
+import json
+import numpy as np
+import torch
+from transformers import AutoModelForCausalLM
+
+model = AutoModelForCausalLM.from_pretrained(
+    "google/gemma-2-2b", torch_dtype=torch.bfloat16
+)
+
+with np.load("eiffel_paris_to_rome/deltas.npz", allow_pickle=False) as z:
+    meta = json.loads(str(z["meta"]))
+    delta = {name: z[name] for name in z.files if name != "meta"}
+
+# w += delta, in place, cast to the weight's own dtype/device — this is
+# exactly circuit_diff/Interp_Gemma's interp_gemma.editing.deltas.apply_delta.
+with torch.no_grad():
+    for name, d in delta.items():
+        w = model.get_parameter(name)
+        w[...] += torch.as_tensor(d, dtype=w.dtype, device=w.device)
+
+# model is now M_e for this case; `meta["request"]` records what was edited.
+```
+
+Applying the same delta a second time double-edits the model; applying it to
+anything other than an unmodified `google/gemma-2-2b` checkpoint does not
+reconstruct a meaningful edited model — `meta["base_model"]` records which
+base the delta was captured against, and callers should assert on it (as
+`interp_gemma.editing.deltas.reconstruct_into` does) before applying.