Ghostgim/cybersec-fact-recall

Cybersec Fact-Recall Benchmark (GhostLM v2)

Free-form short-answer benchmark for small cybersecurity language models. Built and used by the GhostLM project as the truth metric for the ghost-base v1.0 acceptance gate.

Why this exists

Multiple-choice cybersec benchmarks like CTIBench and SecQA reward register matching (the model picks the option that "looks like" a security answer) as much as actual factual recall. A small from- scratch model can hit 28-30% on those without knowing a single CVE identifier. Free-form generation strips that crutch away: if the model knows EternalBlue is CVE-2017-0144, it surfaces the ID; if it doesn't, it confabulates a plausible-looking but wrong CVE.

Dataset structure

100-question seed (v2.0; growing toward 200 in subsequent releases):

Topic	Count
cve	30
mitre	15
cwe	15
protocol	11
owasp	10
crypto	10
tool	6
misc	3

One JSON object per line with fields:

{
  "id":             "fr2-001",
  "topic":          "cve",
  "prompt":         "What is the CVE identifier for the EternalBlue SMB vulnerability?",
  "answer":         "CVE-2017-0144",
  "alternates":     ["CVE-2017-0144"],
  "boundary_match": true,
  "disqualifiers":  ["EternalBlue"],
  "must_appear":    ["RFC 7519", "JWT"]
}

Every record has id, topic, prompt, answer. The rest is optional; defaults are alternates: [], boundary_match: true, disqualifiers: [], must_appear: [].

Grading semantics

Three schema fields tighten the grader vs naive substring match:

• boundary_match (default true). Match only when the candidate appears with non-alphanumeric characters on both sides. "10" no longer matches inside "100".
• disqualifiers. List of phrases that, if present in the completion, void any match credit. Used on echoing-risk questions where the model can surface the answer simply by quoting back the question prompt.
• must_appear. ALL listed substrings must appear (boundary- matched). Used for composite-fact questions like "what RFC defines JWT?" requiring both "RFC 7519" and "JWT" together.

Reference grader: scripts/eval_fact_recall_v2.py in the GhostLM repo (Apache 2.0). 11 unit tests cover the boundary, disqualifier, and must_appear edge cases.

Baseline numbers (2026-05-07)

GhostLM ghost-small (45-81M parameters) chat checkpoints score at the floor across the board. Greedy generation, 120-token budget, M4 MPS.

Checkpoint	Params	Hits / 100	Rate	Per-topic hit
ghost-small v0.4 chat-v3	45M	0	0.0%	(none)
ghost-small v0.7 chat	81M	1	1.0%	misc 1/3
ghost-small v0.9 chat	81M	1	1.0%	owasp 1/10

The line saturates at the bench floor as expected: at this parameter count the model has the register of cybersec writing but not the facts in any retrievable form. Both v0.7 and v0.9 "hits" are likely spurious echoes of question keywords (the misc topic asks about CVSS bounds; the owasp topic asks about Top-10 labels that appear in many corpus passages). The smart grader doesn't invent false positives but it can't distinguish "knows the fact" from "has seen the substring close to similar prose".

Acceptance criterion

The GhostLM project uses >=30% on this bench as one of three ways the next architectural rung (ghost-base, ~360M parameters) can clear its acceptance gate. If a 360M from-scratch model hits 30%+ free-form fact recall on cybersec questions, that validates the parameter-scaling hypothesis. The other two gates are CTIBench full

=40% and the in-repo CTF MCQ >=65%.

Loading

from datasets import load_dataset
ds = load_dataset("Ghostgim/cybersec-fact-recall", data_files="fact_recall_bench_v2.jsonl")
print(ds["train"][0])

Or via the GhostLM grader directly:

git clone https://github.com/joemunene-by/GhostLM
cd GhostLM
PYTHONPATH=. python3 scripts/eval_fact_recall_v2.py \
    --checkpoints my-model=path/to/checkpoint.pt \
    --bench data/raw/fact_recall_bench_v2.jsonl

License

Apache 2.0. Same license as the GhostLM source code.

Citation

@misc{munene2026ghostlm,
  title         = {GhostLM: a from-scratch cybersecurity language model on a transparent scale ladder},
  author        = {Munene, Joe},
  year          = {2026},
  howpublished  = {\url{https://github.com/joemunene-by/GhostLM}},
  note          = {Free-form fact-recall benchmark v2; 100-question seed, growing toward 200; ghost-small line scores 0-1\% (floor)}
}