jsantillana/vectrayx-combined-v2-corpus

VectraYX combined_v2 training corpus

Training corpus of combined_v2, the reference checkpoint of the Qwen2-VL-tower pathway of the VectraYX-Vision-1B paper (weights). Companion to the B8 benchmark.

Contents

• combined_v2_train.jsonl — 11,863 records. combined_v2_heldout.jsonl — held-out split.
• images/ — 6,000 synthetic renders of rendered-UI templates (SIEM dashboard, log table, hex dump, register panel, call stack, MITRE matrix, CVE card, phishing email, gdb and WinDbg sessions, IDA disassembly, Ghidra decompiler, CPU pipeline diagram, Wireshark packets, ANSI terminal, crypto diagram, and the pilot templates).

Each record: messages (a user question and an assistant answer), image_path (relative to this repo), domain, field (which ground-truth field the question asks about, e.g. siem.query), gate (whether the field belongs to a gate), plain_answer (the bare ground-truth value). The assistant answer is <|think|>...</think> followed by the answer; the reasoning text is built deterministically from the renderer's own verified value, with no teacher model, so it cannot contain a fact the image does not show.

Ground truth

Every answer comes from the renderer's own return value at the moment the image was drawn, not from a description written afterwards. This is the property that separates this corpus from the project's earlier LLM-described corpus.

Caveats

• Synthetic renders from one generator family (one font family, fixed palettes and layouts per template). Nothing here supports a claim about real screenshots.
• For most fields the answer alphabet is small enough that every held-out value also occurs in training. The paper reports which fields are exceptions (hexdump.base_addr, logtable.target_row_pid) and one contaminated field (siem.query, 63% of held-out answers occur in training).
• The B8 benchmark is a separate item set; see the paper for the protocol before comparing numbers.

Citation

VectraYX-Vision-1B (paper in preparation).