Fix the byte-identity claim, the placeholder dataset id and the extras description

README.md

@@ -53,7 +53,7 @@ dataset exists for, and it is asserted at build time.
 
 ```
 data/<config>/test-*.parquet   the evaluation set, one row per case
-extras/<config>/               the same arrays in numpy form, plus the GFP tensors
+extras/<config>/               the same arrays as .npy, plus the GFP tensors
   case_index.parquet           case id, edge id, label, typology, HT weight
 ml_baselines/weights/          all three ensemble members, 5 seeds each
 ml_baselines/test_probs/       full-test-split probabilities for the 3 ensemble members
@@ -73,13 +73,13 @@ ml_baselines/test_probs/       full-test-split probabilities for the 3 ensemble
 | `ht_weight` | float64 | Horvitz-Thompson importance weight. Apply when aggregating P, R and F1. |
 | `n_transactions` | int32 | Transactions in the two-hop subgraph. |
 | `n_tokens_approx` | int32 | Approximate prompt length. |
-| `typed_graph_text` | string | The typed-graph prompt, byte-identical to what the evaluated models were given. |
+| `typed_graph_text` | string | The typed-graph prompt the evaluated models were given, identical except for the case identifier token (see Case identifiers). |
 
 ## Usage
 
 ```python
 from datasets import load_dataset
-ds = load_dataset("<org>/amlcompact", "HI-Small", split="test")
+ds = load_dataset("natnitaract/amlcompact-eval", "HI-Small", split="test")
 print(ds[0]["case_id"], ds[0]["label"], ds[0]["typology"])
 print(ds[0]["typed_graph_text"][:200])
 ```
@@ -146,8 +146,8 @@ genuinely null. Rows are in `case_id` order, which is the join key against the r
 predictions and the rubric annotations, so the file is deliberately not sorted to put labelled
 rows first.
 
-To browse labels and weights without pulling the prompt column, use `extras/<config>/`, which
-carries the same arrays in numpy form at 1.6 MB total.
+To browse labels and weights without pulling the prompt column, use `extras/<config>/`: the
+same arrays as `.npy`, plus `case_index.parquet` as a join table, 1.5 MB in total.
 
 ## Case identifiers