---
license: other
task_categories:
  - feature-extraction
  - text-matching
language:
  - code
tags:
  - binary-analysis
  - decompilation
  - ghidra
  - contrastive-learning
  - cross-architecture
  - windows-pe
pretty_name: BSCA Gold v3 Multidomain
---

# BSCA Gold v3 Multidomain

Address-grounded P1 pairs for stripped pseudo-C → source retrieval.

- Dataset ID: `GD_19330e06aae0462447c1fd05ccaa38d7`
- Accepted P1 pairs: 42449
- Repositories: 138
- Target formats: `{"elf": 40733, "pe": 1716}`
- Target architectures: `{"aarch64": 1672, "x86": 1903, "x86_64": 38874}`
- Internal quality GPA: 3.660; target pass: `True`

Use `train.jsonl` for fitting, `development.jsonl` for model selection, and
the immutable `test.jsonl` only after selection.  `dataset_card.json`,
`producer_verification.json`, and `quality_audit.json` are integrity and
scope artifacts, not optional documentation.

Scope: ELF/Linux and PE x86_64 Windows-GNU direct-translation-unit views.
It does not represent Mach-O, MSVC/PDB ABI, raw firmware images, project-build
coverage, or an external blind SCA promotion result.

Repository: `Labradorlabs/bsca-binary-source-gold-v3-multidomain`.