robbd commited on
Commit
ef98458
·
verified ·
1 Parent(s): 0748d57

APPS-541: publish verified current a41 text vectors

Browse files
embeddings/gemma2-768-search-v1-apps542-a41/README.md ADDED
@@ -0,0 +1,61 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Embeddings of the Epstein Index corpus
2
+
3
+ Vectors for the text of the documents that the Epstein Index serves, all in one pinned embedding space. This folder is rebuilt from the served corpus at each release.
4
+
5
+ ## Terms
6
+
7
+ - **Only public, served content.** A vector is published only for a document that the site serves. Anyone could compute the same vector from the same public content.
8
+ - **Held ids are excluded.** No document that touches the holdlist, or wholly duplicates one that does, contributes a vector. The export checks this itself and counts what it refuses (below, zeros included).
9
+ - **Rebuilt each release.** The set follows the site: text that is repaired or withdrawn changes or removes its vectors.
10
+ - **A newer release replaces the set.** Each release replaces this whole folder, so a file from an older release is not valid beside a newer one. Pin the release you used.
11
+
12
+ ## The space
13
+
14
+ | Field | Value |
15
+ | --- | --- |
16
+ | model | `google/embeddinggemma-2` |
17
+ | revision | `914f7f89142e33e77833254d9c9b90c3cef7303b` |
18
+ | dimension | 768 |
19
+ | normalisation | L2, unit length (cosine similarity is the dot product) |
20
+ | document prefix | `title: none | text: ` (no title) before each block |
21
+ | query prefix | `task: search result | query: ` before a query you embed yourself |
22
+ | space label | `embed:google/embeddinggemma-2@914f7f89142e33e77833254d9c9b90c3cef7303b:768:search-v1` |
23
+ | vector | float32[768], fixed-size list |
24
+
25
+ ## The block rule
26
+
27
+ Rule `blocks-v2`: a block is a slice of one served document's cleaned text, packed to 512 tokens and flushed at 768, with a break before a heading or an email `From:` line once the block holds 64 tokens. A block never crosses the boundary of a document; documents are cut by the `hybrid` rule the site uses. Blocks with the same text share one vector (`input_hash`). The block manifest was built 2026-10-15T15:10:54+00:00 (`blocks-v2.parquet` sha256 `19238e2d336fe4a6313934b25ab57246c245dcec4db3bcfabe38f339591ca97b`), and the stored vectors it names hash to `beb08a6e9fe3f1f5a4eba7d31a265f2a06fb7d7b590dff037f1b9e9860c4db0a` (sha256 over each distinct `input_hash` in order, the hash text then its float32 bytes).
28
+
29
+ ## Files
30
+
31
+ | File | Rows | Bytes | sha256 |
32
+ | --- | --- | --- | --- |
33
+ | `documents.parquet` | 43874 | 128186064 | `6582528d2728d0774c988a969a91aad7c1d1beb58a0c34fd4cd568683bd1f40a` |
34
+ | `text/blocks-00000-of-00003.parquet` | 102404 | 296862431 | `f4942f3406ab8f53016913b0a04b940439ea073f840e7ceb2e403b7c30907584` |
35
+ | `text/blocks-00001-of-00003.parquet` | 102404 | 294743020 | `be4344176aa1b17d0280b44de0085f70a2c8985710197ced5c57c1553bd45081` |
36
+ | `text/blocks-00002-of-00003.parquet` | 102404 | 296801955 | `f848c24cfc7139ffc2df2ba2543b58135790f2ab0efc5fe128fed6ef202c3a19` |
37
+
38
+ `README.md` and `manifest.json` carry the same counts and hashes for a machine; the manifest lists `README.md` and not itself.
39
+
40
+ ## Counts
41
+
42
+ | Count | Value |
43
+ | --- | --- |
44
+ | served documents | 48579 |
45
+ | documents in this set | 43874 |
46
+ | blocks in this set | 307212 |
47
+ | distinct block texts (vectors) | 278243 |
48
+ | documents refused, held | 0 |
49
+ | documents refused, not served | 0 |
50
+ | blocks refused, held | 0 |
51
+ | blocks refused, not served | 0 |
52
+
53
+ ## Joining to the documents
54
+
55
+ - `doc_id` is the document id as the site prints it. The page is `https://epstein-index.org/documents/<slug>/` and the text is `https://epstein-index.org/documents/<slug>.txt`, where `<slug>` is the `doc_id` in lower case with its first `:` or `/` replaced by `-` and a trailing `.md` removed (`ds9:EFTA00039025` is `ds9-efta00039025`).
56
+ - A block is a row of `text/blocks-*.parquet`, keyed by (`doc_id`, `ordinal`). `first_page` and `last_page` are the page keys of its first and last page, and `first_page_index` and `last_page_index` count pages in the document from 0. `char_start` and `char_end` are offsets into the document's cleaned text, with its pages joined by a blank line. The served `.txt` can differ from that text by later repairs, so treat the offsets as approximate there.
57
+ - `documents.parquet` has one vector per document: the mean of its block vectors weighted by `content_tokens`, renormalised to unit length. It is the vector that the explore map uses. `blocks` counts the document's blocks.
58
+
59
+ ## Images
60
+
61
+ No image vectors are published in this release (no image set was given or it does not exist).
embeddings/gemma2-768-search-v1-apps542-a41/documents.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6582528d2728d0774c988a969a91aad7c1d1beb58a0c34fd4cd568683bd1f40a
3
+ size 128186064
embeddings/gemma2-768-search-v1-apps542-a41/manifest.json ADDED
@@ -0,0 +1,108 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "blocks_manifest": {
3
+ "built_at": "2026-10-15T15:10:54+00:00",
4
+ "file": "blocks-v2.parquet",
5
+ "inputs": {
6
+ "applied_merges": {
7
+ "sha256": "726fb85c0a2aea4cda062bf8c551bd68d14c1c74f7ef54f924085cbe50f8602c",
8
+ "still_the_input": true
9
+ },
10
+ "documents_v2": {
11
+ "sha256": "7a142f7a871edd5383e47cc98e25a712cd0868847f6f8508b866bc0d6165753a",
12
+ "still_the_input": true
13
+ },
14
+ "holdlist": {
15
+ "sha256": "59f58dca97b30128bc645c06b6885db5f7878eaf8884a9614baa93e51a7b2ac5",
16
+ "still_the_input": true
17
+ },
18
+ "pages": {
19
+ "sha256": "a41ae5a7af46503248db7eb197d583ed205ce2b6c88225a42f79c1bdf09c8ca2",
20
+ "still_the_input": true
21
+ }
22
+ },
23
+ "sha256": "19238e2d336fe4a6313934b25ab57246c245dcec4db3bcfabe38f339591ca97b"
24
+ },
25
+ "documents": {
26
+ "documents": 43874
27
+ },
28
+ "files": [
29
+ {
30
+ "bytes": 4337,
31
+ "path": "README.md",
32
+ "sha256": "b9be3f9d025a7ab8a566e8ee1726f49987994691bc1fbd1193bc74efbdf105ac"
33
+ },
34
+ {
35
+ "bytes": 128186064,
36
+ "path": "documents.parquet",
37
+ "rows": 43874,
38
+ "sha256": "6582528d2728d0774c988a969a91aad7c1d1beb58a0c34fd4cd568683bd1f40a"
39
+ },
40
+ {
41
+ "bytes": 296862431,
42
+ "path": "text/blocks-00000-of-00003.parquet",
43
+ "rows": 102404,
44
+ "sha256": "f4942f3406ab8f53016913b0a04b940439ea073f840e7ceb2e403b7c30907584"
45
+ },
46
+ {
47
+ "bytes": 294743020,
48
+ "path": "text/blocks-00001-of-00003.parquet",
49
+ "rows": 102404,
50
+ "sha256": "be4344176aa1b17d0280b44de0085f70a2c8985710197ced5c57c1553bd45081"
51
+ },
52
+ {
53
+ "bytes": 296801955,
54
+ "path": "text/blocks-00002-of-00003.parquet",
55
+ "rows": 102404,
56
+ "sha256": "f848c24cfc7139ffc2df2ba2543b58135790f2ab0efc5fe128fed6ef202c3a19"
57
+ }
58
+ ],
59
+ "images": {
60
+ "published": false,
61
+ "reason": "no image set was given or it does not exist"
62
+ },
63
+ "rule": {
64
+ "atom_limit_tokens": 768,
65
+ "boundaries": "hybrid",
66
+ "flush_tokens": 768,
67
+ "min_break_tokens": 64,
68
+ "prompt": "document, title none",
69
+ "target_tokens": 512,
70
+ "version": "blocks-v2"
71
+ },
72
+ "schema": "epstein-index.embeddings.export.v1",
73
+ "scope": {
74
+ "held_documents": 0,
75
+ "holdlist_ids": 8130,
76
+ "served_documents": 48579
77
+ },
78
+ "slug": "gemma2-768-search-v1-apps542-a41",
79
+ "space": {
80
+ "dimensions": 768,
81
+ "document_prefix": "title: none | text: ",
82
+ "label": "embed:google/embeddinggemma-2@914f7f89142e33e77833254d9c9b90c3cef7303b:768:search-v1",
83
+ "model": "google/embeddinggemma-2",
84
+ "normalization": "l2-unit",
85
+ "query_prefix": "task: search result | query: ",
86
+ "revision": "914f7f89142e33e77833254d9c9b90c3cef7303b",
87
+ "space": "embed:google/embeddinggemma-2@914f7f89142e33e77833254d9c9b90c3cef7303b:768:search-v1",
88
+ "vector": "float32[768], fixed-size list"
89
+ },
90
+ "text": {
91
+ "blocks": 307212,
92
+ "distinct_block_texts": 278243,
93
+ "documents": 43874,
94
+ "manifest_blocks": 307212,
95
+ "manifest_documents": 43874,
96
+ "refused": {
97
+ "blocks": {
98
+ "held": 0,
99
+ "not_served": 0
100
+ },
101
+ "documents": {
102
+ "held": 0,
103
+ "not_served": 0
104
+ }
105
+ },
106
+ "store_content_sha256": "beb08a6e9fe3f1f5a4eba7d31a265f2a06fb7d7b590dff037f1b9e9860c4db0a"
107
+ }
108
+ }
embeddings/gemma2-768-search-v1-apps542-a41/text/blocks-00000-of-00003.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f4942f3406ab8f53016913b0a04b940439ea073f840e7ceb2e403b7c30907584
3
+ size 296862431
embeddings/gemma2-768-search-v1-apps542-a41/text/blocks-00001-of-00003.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:be4344176aa1b17d0280b44de0085f70a2c8985710197ced5c57c1553bd45081
3
+ size 294743020
embeddings/gemma2-768-search-v1-apps542-a41/text/blocks-00002-of-00003.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f848c24cfc7139ffc2df2ba2543b58135790f2ab0efc5fe128fed6ef202c3a19
3
+ size 296801955