Measured on a 4-vCPU VM with pre-downloaded shards: 50 MB/s of JSONL output (12 MB/s per worker process; CPU-bound).
One input shard (355 MB parquet) yields ~600 MB JSONL (75k records) in 30 s per worker; shard download from the Hub measured at ~35 MB/s per stream (10 s per shard), so download roughly matches processing when run with --workers 4.
Extrapolation: 40 GB JSONL ~= 67 shards -> ~13 min of processing at 50 MB/s (plus ~11 min of downloads if not overlapped; the script overlaps them per worker). On 8 cores expect roughly half. Parquet conversion (zstd) adds ~2x the generation time and shrinks size ~2.3x.
Full 880-shard corpus (permissive subset) would yield ~598 GB JSONL.
Clusters (cross-document coreference)
Clusters by kind (repo = real multi-file repo, bundle = synthetic project of unrelated files, singleton = one file):
cluster_kind
count
%
singleton
1,161,632
48.8
repo
612,711
25.8
bundle
604,968
25.4
Records by cluster kind:
cluster_kind
count
%
bundle
2,157,603
40.9
repo
1,959,165
37.1
singleton
1,161,632
22.0
Files per cluster:
cluster_size
count
%
1
1,161,785
48.8
2
562,133
23.6
3
249,362
10.5
4
135,923
5.7
5
94,680
4.0
6
65,780
2.8
7
44,693
1.9
8
64,955
2.7
Entities
entity_type
count
%
api_key
1,384,896
35.4
person
717,813
18.3
password
674,978
17.2
token
498,684
12.7
email
341,310
8.7
private_key
296,867
7.6
Documents per entity (all clusters; 1 = no cross-document recurrence):
n_docs
count
%
1
2,747,309
70.2
2
807,464
20.6
3
249,818
6.4
4
80,686
2.1
5
23,140
0.6
6
6,131
0.2
Mentions per entity:
n_mentions (8 = 8+)
count
%
1
1,990,693
50.9
2
933,660
23.9
3
457,212
11.7
4
247,281
6.3
5
147,967
3.8
6
66,018
1.7
7
33,686
0.9
8
38,031
1.0
Secret spans by surface form type:
surface_form_type
count
%
literal
6,509,816
82.0
full_name
435,636
5.5
username
354,633
4.5
escaped_literal
255,658
3.2
email
217,690
2.7
first_name
108,611
1.4
last_name
32,587
0.4
handle
21,739
0.3
Secret spans by label
label
count
%
api_key
2,939,108
37.0
password
1,424,697
18.0
token
1,045,257
13.2
email
943,133
11.9
private_key
630,969
8.0
person_name
576,834
7.3
username
376,372
4.7
Secret spans by value_type
value_type
count
%
generic_hex_key
2,087,008
26.3
password
1,424,697
18.0
email
943,133
11.9
pem_private_key
630,969
8.0
full_name
435,636
5.5
username
354,633
4.5
generic_hex_token
180,240
2.3
aws_access_key_id
139,245
1.8
generic_alnum_token
119,300
1.5
basic_auth_b64
109,577
1.4
first_name
108,611
1.4
generic_dashed_token
102,788
1.3
bearer_token_opaque
95,163
1.2
jwt
91,667
1.2
generic_lower_key
85,773
1.1
generic_base64_secret
85,494
1.1
generic_hex_key_upper
84,085
1.1
generic_alnum_key
68,842
0.9
github_pat
68,772
0.9
aws_secret_access_key
67,916
0.9
digitalocean_pat
50,695
0.6
google_api_key
46,089
0.6
slack_token
45,792
0.6
stripe_secret_key
45,511
0.6
generic_upper_key
42,887
0.5
generic_upper_token
40,271
0.5
last_name
32,587
0.4
mailgun_api_key
30,031
0.4
generic_hex_dashed_key
29,657
0.4
twilio_api_key
28,909
0.4
gitlab_pat
26,254
0.3
discord_bot_token
23,618
0.3
npm_token
22,979
0.3
hf_token
22,879
0.3
slack_webhook
22,810
0.3
github_fine_grained_pat
22,452
0.3
handle
21,739
0.3
azure_client_secret
21,648
0.3
sendgrid_api_key
20,612
0.3
anthropic_api_key
20,539
0.3
openai_api_key
20,256
0.3
generic_digit_key
14,606
0.2
Hard-negative spans by value_type
value_type
count
%
ssh_public_key
414,055
8.4
uuid
413,392
8.4
example_email
412,768
8.3
placeholder_name
412,687
8.3
base64_non_secret
412,630
8.3
sha1_git_sha
412,551
8.3
public_key_pem
412,511
8.3
env_var_reference
412,495
8.3
sha256_hash
412,328
8.3
md5_hash
411,959
8.3
version_or_id
411,726
8.3
placeholder
411,582
8.3
Secret spans by syntactic context
context
count
%
assignment
3,567,362
44.9
comment
1,057,198
13.3
json_key
860,567
10.8
yaml_key
810,404
10.2
pem_block
646,725
8.1
auth_header
457,619
5.8
url_credential
277,794
3.5
function_arg
133,346
1.7
env_assignment
125,355
1.6
Records by language
language
count
%
Java
1,013,525
19.2
JavaScript
611,284
11.6
HTML
569,348
10.8
PHP
447,622
8.5
C
418,772
7.9
Markdown
377,703
7.2
C#
374,456
7.1
Python
342,346
6.5
C++
276,426
5.2
Ruby
244,789
4.6
GO
142,183
2.7
TypeScript
118,371
2.2
CSS
77,104
1.5
Scala
53,472
1.0
Shell
52,401
1.0
SQL
19,549
0.4
Haskell
19,013
0.4
Makefile
18,329
0.3
Rust
18,294
0.3
Lua
14,602
0.3
Perl
14,562
0.3
Dockerfile
9,249
0.2
PowerShell
8,941
0.2
TeX
8,167
0.2
CMake
7,311
0.1
Visual Basic
5,753
0.1
Batchfile
5,395
0.1
FORTRAN
3,693
0.1
Julia
3,596
0.1
Assembly
2,144
0.0
File size distribution (text field, sample of 2 shards)
bucket
records
<1 KB
29,976
1-10 KB
99,191
10-100 KB
21,311
>=100 KB
858
Validation (2 shards, 151,336 records)
Offset assertion text[start:end] == value on every span and hard negative: 0 failures (also asserted at generation time for all records).
Regex scanner (gitleaks-style patterns for 15 value types) finds 100% of injected spans of those types at the labeled offsets (secrets are not mangled by quoting/escaping/indentation).
20 random records eyeballed; see validate.py --n-eyeball 20.
Held-out eval (eval_creddata.jsonl)
67,896 real labeled lines from Samsung/CredData (+-3 lines context): 15,753 GroundTruth=T, 48,061 F (hard negatives), 4,082 X (ambiguous/unclear, no span).
