Downloads · 30 days
0
asatheesh/latent-mas-aligner-qwen3-14b-seed-1
latent-mas-aligner-qwen3-14b-seed-1 is a machine learning model from asatheesh. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for pytorch. The card lists the license as apache-2.0.
This is the largest backbone in the LatentShield study. Qwen3-14B's hidden size is 5120, so the input projection is wider than in the 4B aligners (37.8M trainable parameters against 27.3M). Otherwise it is the same de…
Downloads · 30 days
0
Access
Public
Updated Nov 5, 2026
Repo size
151 MB
Likes
0
Public
Click a slice to open those files.
.pt151 MB · 100%
From the Hugging Face model README
This is the largest backbone in the LatentShield study. Qwen3-14B's hidden size is 5120, so the input projection is wider than in the 4B aligners (37.8M trainable parameters against 27.3M). Otherwise it is the same design: attention-pool the pre-Judger latents of a LatentMAS run, project them into Llama-Guard-3-8B's residual stream, and let the frozen guard tail decide. A positive call injects a safety notice for the Judger.
What changes at 14B. Undefended, the 14B LatentMAS completes 29.4% of HarmBench requests. With LatentShield at
p10, HarmBench ASR is 0.050–0.056 across seeds. StrongREJECT moves less (0.121 → 0.105–0.128), and over-refusal rises from 0.025 to 0.055–0.118 on OR-Bench-Hard.
p10| HarmBench ASR ↓ | StrongREJECT ASR ↓ | OR-Bench-Hard RR ↓ | PHTest FRR ↓ | mean ↓ | |
|---|---|---|---|---|---|
| LatentMAS, no defense | 0.294 | 0.121 | 0.025 | 0.009 | 0.112 |
this checkpoint (p10) | 0.050 | 0.115 | 0.055 | 0.068 | 0.072 |
seed 0 (p10) | 0.050 | 0.105 | 0.078 | 0.088 | 0.080 |
seed 2 (p10) | 0.056 | 0.128 | 0.118 | 0.118 | 0.105 |
Of the three Qwen3-14B seeds, this one has the lowest mean failure rate at p10 (0.072; the seeds span 0.072–0.105). Seed-to-seed spread is part of the result, so report all three rather than picking one.
p01 | p05 | p10 | p20 | p50 | |
|---|---|---|---|---|---|
tau | 0.9641 | 0.8775 | 0.7249 | 0.1176 | 0.0000 |
Each is a quantile of this seed's p_unsafe on benign validation rollouts (p05 is
exceeded by 5% of them). The values differ between seeds and backbones, so use the ones on
the card of the checkpoint you load.
[B, 24, 5120] = Planner, Critic and Refiner × 8 latent steps.aligner.pt (151 MB), a dict with d_a, d_g, hidden_dim, pool and state_dict.The frozen Llama-Guard-3-8B tail (decoder blocks 17–31, final norm, LM head; 7.6 GB) is not included in this repo. Build it once from meta-llama/Llama-Guard-3-8B (a gated model; request access first) with python -m scripts.aligner.extract_cli --mode dump-head-bundle --layer 16 --out guard_head_bundle.pt.
from scripts.aligner.aligner_loader import build_aligner_from_bundle
import torch
aligner, arch, meta = build_aligner_from_bundle(torch.load("aligner.pt", weights_only=False))
The full scoring path (guard tail + rotary embedding) is in
scripts/aligner/compute_tau_fpr05.py in the code repo.
Research use only. Requires Qwen3-14B LatentMAS latents. Collection: LatentShield