jsantillana/vectrayx-b8-benchmark

B8 — ground-truth-by-construction extraction gate

Companion benchmark for the VectraYX-Vision-1B paper (vectrayx-vision-1b). B8 is a wider, more rigorously controlled successor to the paper's original nine-field gate (B6/B7 remain separate, tool-identification benchmarks).

What it is

• 2,040 items across 34 fields spanning 16 rendered-UI templates (SIEM dashboards, log tables, hex dumps, register panels, call stacks, MITRE matrices, CVE cards, phishing emails, gdb/WinDbg sessions, IDA disassembly, Ghidra decompiler, CPU pipeline diagrams, Wireshark packet captures, ANSI terminal sessions, crypto primitive diagrams).
• n=60 per field, single seed.
• Every item's ground truth is the renderer's own return value — not an answer written separately about the image, which can silently disagree with what's actually drawn. Every item also carries a fixed within-template shuffle_partner_image_path for a shuffled-image control.

Methodology

Each field is scored against two controls, not one:

• acc_shuffled — same question, a different (wrong) image from the same template. Tests whether the answer depends on this image.
• acc_best_constant — accuracy of always emitting the field's single most common ground-truth value. Tests whether the answer needs to depend on any image at all.

Pre-registered decision rule (frozen before scoring): a field passes iff acc_real >= 0.60 AND acc_real - max(acc_shuffled, acc_best_constant) >= 0.40.

Files

• b8_items.jsonl — the 2,040 items: {item_id, template, field, tier, question, gt, image_path, shuffle_partner_image_path}.
• images/ — 960 rendered PNGs referenced by the items above.
• b8_results_full.jsonl — full model outputs from the domain_expansion_v1 checkpoint reported in the paper (pred_real, pred_shuf per item).
• score_b8_results.py — per-field scorer (normalize + exact-match + char-sim, reproduces the paper's Table tab:b8 from b8_results_full.jsonl).
• build_b8_benchmark.py — the corpus builder (CPU-only, no model needed; regenerates images/ + b8_items.jsonl from the renderer source, which lives in the main vectrayx-vision-1b code release).

Headline result

9 of 34 fields clear the gate (shuffled- and best-constant-confirmed genuine grounding). The other 25 split into three explained categories — see the paper (\S\ref{sec:results:b8}) for the full breakdown and the accompanying "landmark lookup, not free-text reading" discussion.

Added 2026-09-29: outputs for a second checkpoint (combined_v2)

• b8_results_combined_v2.jsonl — per-item model outputs (real and shuffled-image conditions) for combined_v2, the reference checkpoint trained from scratch on all 16 templates (weights, training corpus). The paper reports 12 of 34 fields passing under the scorer used for this run; a later re-implementation of the scorer from the field specification gives 14 of 34, and the two counts have not been reconciled.
• b6_b7_combined_v2.json — the B6/B7 result file for the same checkpoint (tool identification 0.000 on all 50 items).