Datasets:
license: other
task_categories:
- text-generation
configs:
- config_name: default
data_files:
- split: train
path: train.jsonl
- split: validation
path: validation.jsonl
- split: test
path: test.jsonl
language:
- en
- ru
tags:
- cybersecurity
- malware
- malware-research
VX direct pairs v2
VX corpus coverage
The source collection recursively traversed these VX Underground roots:
/Papers, including its subdirectories./Malware Analysis/2023./Malware Analysis/2024./Malware Analysis/2025./Malware Analysis/2026.
This is a collected snapshot, not a continuously synchronized mirror or a claim that every file currently available on the website was successfully included. The years above apply to the Malware Analysis roots; Papers was not restricted to publications from 2023 onward.
The collection included associated reports, papers and sample/artifact folders. The direct-pair pipeline used successfully ingested readable material, including documents, source code, YARA rules and readable archive members. Related artifacts were grouped with their reports where that relationship was available. Opaque executables and binary blobs were represented by inventory/static metadata, not executed or treated as observed sandbox behavior. Unsupported, unreadable or unresolved archive content was not equivalent to extracted teaching evidence.
Security-scope filtering excluded incidental bundled vendor libraries and unrelated administrative/financial material from generation, without discarding security-relevant code or reports solely because of a directory name. These filters, extraction failures, deduplication and generation/review decisions mean the final pairs are a selected derivative of the collection, not one example per downloaded file or exhaustive coverage of every source context.
Portable messages JSONL: 173,325 records. Sources are combined once at natural sizes, without weighting or oversampling. Exact message duplicates are removed, first source wins, and identical learner prompts stay in one split (test takes precedence over validation, then train). No content is truncated.
VX material is best-effort synthetic direct question/answer data, mostly unreviewed; residual factual errors remain possible. Merge validation is structural, not a semantic quality approval. WhiteRabbit content, when present, is preserved using the explicit mapping in mix.yaml; a dataset name does not certify ReAct behavior. Question/response mapping excludes upstream system text and other columns.
Use load_dataset("json", data_files={"train": "train.jsonl", "validation": "validation.jsonl", "test": "test.jsonl"}) locally, or load the pinned Hub revision. Apply the student's matching tokenizer/chat template and assistant-only loss downstream. Token-length preflight is required before training; this package neither tokenizes nor silently truncates examples.
See release.json for exact source revisions, hashes, split counts, duplicate count and split moves; origins.jsonl links each output row to its source. The sources/ directory preserves upstream cards and license/notice files. VX source excerpts have heterogeneous third-party rights: no blanket redistribution license is asserted. WhiteRabbit's individual licenses/usage conditions remain applicable; this aggregate is not entirely Apache-2.0. See docs/dataset-merging.md in the docprep repository for the preparation, verification and training workflow.