Modelpublic

MuVeraAI/guard-lora-t1c-mlp

sourceHugging Faceotherupdated 24d agoView on Hugging Face

0likes16downloads

Model Card

guard-lora-t1c-mlp — a guard adapter that lost an output class

LoRA adapter on `Qwen/Qwen3Guard-Gen-0.6B`, target modules: MLP only.

Feed-forward projections only. Disjoint from arm B, and independently sufficient to destroy the class.

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

Read this before using it

This adapter emits only two severity values. The base model emits three — Safe, Unsafe, Controversial. This one never emits Controversial.

	recall	FPR	`safe`	`unsafe`	`controversial`
Qwen3Guard-Gen-0.6B (base)	0.7150	0.0854	1290	863	122
this adapter	0.7930	0.0864	1191	1084	0

ExpGuardTest, n = 2,275, `num_ctx` 8192, seed 0, greedy. Paired exact-binomial McNemar against the base: 19 items the base caught and this missed, 117 the reverse, p < 0.0001.

Consequence for anyone deploying it: a policy control of the form Controversial = SAFE becomes a silent no-op. It will read as configured and govern nothing. If your stack has such a lever, this adapter disables it without saying so.

Recall is not the whole story. Higher recall here coexists with a lost output class, and for the attention-only arm with a false-positive rate that nearly doubles the base. A model scoring better on the headline metric is not the same instrument.

Why it was published

It is a negative result, and the useful kind. Three arms were trained, varying only which projections the adapter touches. All three destroyed the class. Attention-only and MLP-only are disjoint families and each is independently sufficient, so the loss is not attributable to a module family. The cause is the training target: the corpus renders the Safety: field as a boolean, so ~11,272 gradient updates teach a two-valued field.

—`guard-lora-t1a-all-modules` — control, all seven
—`guard-lora-t1b-attention` — attention only
—`guard-lora-t1c-mlp` — MLP only

Per-item verdicts for every run: `MuVeraAI/guard-verdicts`.

Training


base	`Qwen/Qwen3Guard-Gen-0.6B` (Apache-2.0)
method	LoRA, rank 16, alpha 32, one epoch, bf16
target modules	`gate_proj, up_proj, down_proj`
corpus	11,272 rows derived from ExpGuardMix (`6rightjade/expguardmix`)
hardware	single A100-80GB

Licence — research only, and narrower than the base model

The base model is Apache-2.0. This adapter is not.

It was trained on data derived from ExpGuardMix, whose access gate requires affirming research-only use — a restriction narrower than that dataset's own CC-BY-4.0 licence. These weights are a derivative of that data, so the restriction is passed forward rather than dropped:

Research and evaluation use only. Not for commercial use. If you intend to use these weights commercially, obtain your own clearance for ExpGuardMix-derived artifacts first.

The training corpus itself is not redistributed here or anywhere else in this release.

Citation

bibtex

@misc{jha2026guardlorat1cmlp,
  author = {Vikram Jha},
  title  = {guard-lora-t1c-mlp: a guard adapter that lost an output class},
  year   = {2026},
  note   = {ORCID 0009-0004-3959-6099},
  howpublished = {\url{https://huggingface.co/MuVeraAI/guard-lora-t1c-mlp}}
}