GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
jhk0317
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
jhk0317
Model Tree
Qwen/Qwen3.8-27B
Input Modalities
Output Modalities
Supported Functionality
An earlier (2026-09-16) release of this study picked the ablation layer/alpha using 512-token,
substring-matched refusal scoring (is_refusal() on generated text). That metric overcounts
bypass: 64-256 token generations often get cut off mid-hedge or mid-redirect, which the substring
matcher misreads as a completed harmful compliance.
This repository is from a 2026-09-26 re-run that instead uses 2048-token generations scored
by an LLM-judge ensemble (gpt-5.6-terra + claude-haiku-4-5-20251001, mean of both, with
human-adjudicated labels for judge-rejected calls). The layer selection lands on the same layer
(L46) as before, but the honest bypass/refusal rates at each alpha are substantially different from
the original release's substring-based numbers — use the numbers on this card, not any substring
based ones you may see elsewhere for this study.
| α | harmful-prompt refusal rate (judge) | harmless-prompt refusal rate (judge) | KL vs. stock Qwen3.8-27B |
|---|---|---|---|
| 0.25 | 1.00 | 0.00 | 0.0025 |
| 0.5 | 0.98 | 0.00 | 0.013 |
| 0.75 | 0.94 | 0.00 | 0.035 |
| 1.0 (this repo) | 0.55 (partial compliance 0.1) | 0.00 | 0.095 |
This repo is the α=1.0 arm: harmful-prompt refusal rate 0.55 (partial compliance 0.10), harmless-prompt refusal rate 0.00, KL-vs-stock 0.095.
Two things worth being explicit about:
Apache-2.0, inherited from the base model. See LICENSE and NOTICE.md in this repo — this is a
modified derivative of Qwen/Qwen3.8-27B (directional ablation, not fine-tuning).
Interpretability and AI-safety research into how refusal behavior is represented and how robust it is to a simple linear intervention. This is a research artifact, not a general-purpose assistant release, and it has not been safety-tuned after the edit.