MuVeraAI/guard-verdicts

Guard Model Verdicts

27,818 per-item guard verdicts across 14 evaluation runs, 10 model variants and 2 corpora.

Guard-model papers report aggregate recall. This dataset publishes the layer underneath: what each model said about each item, so that a reader can recompute the aggregates rather than trust them.

Author: Vikram Jha, MuVeraAI · ORCID 0009-0004-3959-6099 · vikram@muveraai.com

What is NOT here, and why that matters

This dataset contains no prompt text. Rows carry a positional sample_id (egt-00000, wgt-00000), the ground-truth label, the model's verdict, its severity value and categories — nothing from the underlying corpora.

That is deliberate. The corpora are ExpGuardMix (6rightjade/expguardmix) and WildGuardMix (allenai/wildguardmix), both gated. ExpGuardMix's gate requires affirming research-only use, which is narrower than its CC-BY-4.0 licence. Redistributing the items would route around that gate.

Joining back is easy for anyone who should be able to. sample_id is the row's position in the corpus's own test split, sorted by id. Accept the upstream gate, load the split, and the join is positional. Nobody who has not accepted the gate gets corpus content from this dataset.

Why the severity column is the interesting one

Every fine-tune in this programme damaged its severity field, and the damage is invisible to a recall-only review.

model	recall	FPR	controversial verdicts
Qwen3Guard-Gen-0.6B (base)	0.7150	0.0854	122
T1-A — LoRA, all 7 target modules	0.7842	0.0589	0
T1-B — LoRA, attention only	0.8439	0.1649	0
T1-C — LoRA, MLP only	0.7930	0.0864	0

ExpGuardTest, n = 2,275, num_ctx 8192, seed 0, greedy.

Qwen3Guard emits a three-valued severity: Safe, Unsafe, Controversial. Every fine-tune here emits two. T1-B has the highest recall in the programme and is the worst model in it — the gain is substantially bought with false positives, and it has lost an output class. A Controversial = SAFE policy control becomes a silent no-op on any of these models.

Attention-only and MLP-only are disjoint module families and each independently destroys the class, so the loss is not attributable to a module family. The cause is the training target: the corpus renders Safety: as a boolean, and 11,272 gradient updates teach a two-valued field.

Files

file	contents
verdicts.jsonl	27,818 rows, one per (run, item). The flat table.
runs.json	Per-run metadata: metrics, sampling options, result_digest, schema version.
runs/*.json	The complete original result documents, verbatim.

Fields: run_id, corpus, model, sample_id, expected_unsafe, predicted_unsafe, severity, categories, gated_by_category, errored.

errored matters. The harness fails closed — a backend failure is scored as harmful — so an errored row inflates recall. Exclude these rows before any paired comparison; the papers do.

Reproducing the headline comparisons

from datasets import load_dataset
import itertools, math

d = load_dataset("MuVeraAI/guard-verdicts", split="train")
base = {r["sample_id"]: r for r in d if r["run_id"].startswith("baseline-0.6b")
        and r["corpus"] == "ExpGuardTest"}
cand = {r["sample_id"]: r for r in d if r["run_id"].startswith("T1-A")}

b_only = c_only = 0
for sid, b in base.items():
    c = cand.get(sid)
    if not c or not b["expected_unsafe"] or b["errored"] or c["errored"]:
        continue
    if b["predicted_unsafe"] and not c["predicted_unsafe"]: b_only += 1
    elif c["predicted_unsafe"] and not b["predicted_unsafe"]: c_only += 1

n, k = b_only + c_only, min(b_only, c_only)
p = min(1.0, 2 * sum(math.comb(n, i) for i in range(k + 1)) / 2**n)
print(b_only, c_only, p)        # 31 116 <0.0001

Paired exact-binomial McNemar, positives only, errored items excluded — the comparison the papers run. Chi-square is not used: several comparisons here have fewer than 25 discordant pairs.

Licence

CC BY 4.0 — these are original measurements. The corpora they were measured on carry their own licences and gates, which this dataset does not alter and does not redistribute.

Citation

@misc{jha2026guardverdicts,
  author = {Vikram Jha},
  title  = {Guard Model Verdicts: per-item results for Qwen3Guard baselines and LoRA fine-tunes},
  year   = {2026},
  note   = {ORCID 0009-0004-3959-6099},
  howpublished = {\url{https://huggingface.co/datasets/MuVeraAI/guard-verdicts}}
}