Replace bfloat16 ground truth with fp32 run; archive v1 under queries/deprecated

README.md

@@ -137,9 +137,7 @@ This verifies recovered query text against `_checksums.json`; it does not compar
 
 ##### Numerical Reproducibility
 
-The original `hit_scores` and `hit_fineweb_ids` were computed using bfloat16 GPU arithmetic, while the regeneration scripts produce float32 embeddings for reproducibility. The resulting numerical differences are small, but may reorder tied results or change membership near the top-1000 cutoff.
-
-We are working on a fix and expect to release it within the next week.
+Supernova searches the float16 FineWeb embeddings using float32 queries. Although ties are resolved deterministically, floating-point arithmetic is not fully deterministic across implementations and hardware, so small numerical differences may perturb result ordering, particularly near the (k=1000) boundary.
 
 ### Intended Uses
 

queries/deprecated/README.md

@@ -0,0 +1,8 @@
+# Deprecated Ground Truth (bfloat16)
+
+These are the original ground-truth files, published when the dataset first shipped and replaced on September 18, 2026. The current files live in the parent `queries/` directory. Use those instead unless you are reproducing a number computed against this version.
+
+The replacement was made because these `hit_scores` were computed with bfloat16 GPU arithmetic. With 10 billion documents, many candidates sit inside that margin, and the scores here are inflated by a large enough amount to reorder the top of the list. The replacement run scores the same corpus with float32 query vectors against the float16 stored corpus using Supernova, with TF32 disabled.
+
+
+Licensing is unchanged. See the dataset README for the FineWeb and MS MARCO terms.

queries/deprecated/_checksums.json

@@ -0,0 +1,18 @@
+{
+  "dense": {
+    "rows": 100000,
+    "query_text_sha256": "3a354db0c5aa2bcf5d4563844c645d8fc23ff1780f83cce28900b7fc917a5756"
+  },
+  "sparse": {
+    "rows": 10000,
+    "query_text_sha256": "fd33c5fe47600b11fe12b7438a342cfa866e2ada82f911726e91c04ad079dfc5"
+  },
+  "text_filters": {
+    "rows": 4953,
+    "query_text_sha256": "80d4a77cc3cbed241399fe01525e07987aa886178dcccf943a070a4e2e391fed"
+  },
+  "structured_filters": {
+    "rows": 5000,
+    "query_text_sha256": "9023ccb0e1afdda88c20a2d83923102d53b33b989b9ce0f2b11a256a16ae9e83"
+  }
+}

queries/deprecated/_sets.json

@@ -0,0 +1,63 @@
+[
+  {
+    "set": "dense",
+    "file": "gt_dense_k1000.parquet",
+    "rows": 100000,
+    "columns": [
+      "msmarco_config",
+      "msmarco_split",
+      "msmarco_query_id",
+      "hit_scores",
+      "hit_fineweb_ids"
+    ],
+    "regenerate_vectors": "dense"
+  },
+  {
+    "set": "sparse",
+    "file": "gt_sparse_k1000.parquet",
+    "rows": 10000,
+    "columns": [
+      "msmarco_config",
+      "msmarco_split",
+      "msmarco_query_id",
+      "hit_scores",
+      "hit_fineweb_ids"
+    ],
+    "regenerate_vectors": "sparse"
+  },
+  {
+    "set": "text_filters",
+    "file": "gt_text_filters_k1000.parquet",
+    "rows": 4953,
+    "columns": [
+      "msmarco_config",
+      "msmarco_split",
+      "msmarco_query_id",
+      "hit_scores",
+      "hit_fineweb_ids",
+      "selectivity_tier",
+      "keyword_phrase",
+      "domains"
+    ],
+    "regenerate_vectors": "dense"
+  },
+  {
+    "set": "structured_filters",
+    "file": "gt_structured_filters_k1000.parquet",
+    "rows": 5000,
+    "columns": [
+      "msmarco_config",
+      "msmarco_split",
+      "msmarco_query_id",
+      "hit_scores",
+      "hit_fineweb_ids",
+      "selectivity_tier",
+      "structured_group",
+      "ls_gte",
+      "date_gte",
+      "date_lt",
+      "dump_set"
+    ],
+    "regenerate_vectors": "dense"
+  }
+]

queries/deprecated/gt_dense_k1000.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0f69daadce2c36f76ee5a2bbfc3d94397a34115ad1685869ae32d8b8db7e1269
+size 2129786818

queries/deprecated/gt_sparse_k1000.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:48e9bc62a0af8c0e50c5aa560763e5ed4106d8bd315e761fdb39ea07a3f36863
+size 214140613

queries/deprecated/gt_structured_filters_k1000.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0f1ac2bfbda019e067b03ff8b1c3e535eaa8717ef92bec6ed6158b62e1afcb70
+size 110162713

queries/deprecated/gt_text_filters_k1000.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:56e3939087a7f2c5b7a450c5daf82dfe02fba903ea6671535a2fd974ddcc8a6b
+size 98190990

queries/gt_dense_k1000.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:0f69daadce2c36f76ee5a2bbfc3d94397a34115ad1685869ae32d8b8db7e1269
-size 2129786818
+oid sha256:847989d235ecd8c7118b64709cfe5d54ff36b82ad8c70c9de31c5347633bbd81
+size 2126961576

queries/gt_sparse_k1000.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:48e9bc62a0af8c0e50c5aa560763e5ed4106d8bd315e761fdb39ea07a3f36863
-size 214140613
+oid sha256:5b1996051eca568e29f5c47f17fd7eb4521154b9224d08c94cf04b322e32ec40
+size 212728905

queries/gt_structured_filters_k1000.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:0f1ac2bfbda019e067b03ff8b1c3e535eaa8717ef92bec6ed6158b62e1afcb70
-size 110162713
+oid sha256:77e4ce1929aacb74ced5056dedfb331ec9ae79884efa0f5d4012014833ecad92
+size 110078476

queries/gt_text_filters_k1000.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:56e3939087a7f2c5b7a450c5daf82dfe02fba903ea6671535a2fd974ddcc8a6b
-size 98190990
+oid sha256:14ca848475bad3e7500d5ce9ce24caf6d87ad1ea360ad50cad66330a4e7f7e18
+size 98007085