GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
yavuz-ai
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
yavuz-ai
Model Tree
Qwen/Qwen3-4B-Instruct-2507
Input Modalities
Output Modalities
Supported Functionality
On a two-axis eval this stage raises over-refusal on benign prompts from 0.327 to 0.431 while safety stays ~1.0. It is published for reproducibility and failure analysis, not as a general-purpose model. See the repo for the stage-by-stage teardown of the harmless-vs-overrefusal tension.
LoRA SFT (r=16, alpha=32, all-linear), completion-only loss on the revised answers, 2 epochs, lr 1e-4, bf16, on 1x A100. Data: ~4,800 self-revisions generated from hh-rlhf harmless-base prompts.