Buckets:

|
download
raw
3.59 kB
# candidate_bucket
Everything we drafted for SecPat before human verification. Claude Code drafted a
role mapping for 502 (repository, pattern) pairs over 392 repositories (200 Python
and 192 Java), and our traversal rules derived an interaction sequence for each draft
from a CPG of its repository. The 281 drafts with at least one interaction are the
qualified records, from which we sampled the 100 records of the companion bucket
`hf://buckets/AnonymousSpottedToad/human_verified_bucket`.
```
candidate_bucket/
├── manifest.csv ··································· 502 (repository, pattern) pairs
├── codebase/<language>/<repo>/ ··································· 392 repositories
├── cpg/ ·································································· 392 CPGs
│ └── <repo>__<JAVASRC|PYTHONSRC>.cpg
├── preliminary/ ············································· 281 qualified records
│ ├── mappings/ ······················································ 1,976 roles
│ │ └── <language>/<pattern>/<repo>.json
│ └── interactions/ ············································· 724 interactions
│ └── <language>/<pattern>/<repo>.json
├── preliminary_without_sequence/ ··························· 221 pairs, same layout
└── transcripts/ ·········································· 1 session, 502 subagents
```
## How the drafts were made
1. Our mining tool used lexical code search to group 400 Python and 347 Java
repositories by the five patterns:
- P1 `authentication_pwd`
- P2 `verifiable_token_based_authentication`
- P3 `obscure_token_based_access_control`
- P4 `session_based_access_control`
- P5 `limit_request_rate`
2. Claude Code (Opus 5.5) spawned one subagent per (repository, pattern) pair. Each
subagent mapped every role of the pattern to a code symbol and marked it present,
external or missing.
3. Joern 4.0.627 built one CPG per repository, and our traversal rules derived the
interactions between the mapped roles from it.
## Read before use
- **Unverified.** On the 100 verified records, the annotators changed 22.8% of the
drafted roles, so expect errors of that order here.
- **Qualified records.** `manifest.csv` marks each pair `qualified`. For a pair that
isn't, `no_sequence_reason` is one of:
- no role in code (118)
- one role in code (5)
- no interaction found (79)
- traversal timeout (19), where the traversal ran past its budget of 870 s
- **Two CPG builds.** We built 378 CPGs through our Joern service, which runs Lombok
processing for Java. We built the last 14 directly with `joern-parse`. The 8 Java
CPGs among those 14 skip Lombok processing, so the methods that Lombok generates are
missing from them. Only `lbk-ones_easy4j` uses Lombok widely, which may affect its
P2 draft.
- **Field names.** These follow `human_verified_bucket`: `role`, `symbol`,
`interaction_sequence` and `order`.
- **Code licenses.** `codebase/` keeps source files only. `manifest.csv` lists the
license that GitHub reports for each repository, and 51 repositories report none.
## License
MIT for this dataset's own content. Each codebase keeps the license of its repository.

Xet Storage Details

Size:
3.59 kB
·
Xet hash:
f56d0ecf8b3c3157ab8162faaaad0d9dc6224660bcf34d9023e0f605e5d303d6

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.