GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Aditi20
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
Aditi20
Model Tree
Qwen/Qwen3-0.6B-Base
Input Modalities
Output Modalities
Supported Functionality
text
surface: tool_call | tool_outputtask: <the user's request>prior steps: <earlier agent actions, or (none)>content:<the proposed command / the tool output>Overall bypass on a held-out red-team set is 0.37 (vs 1.6% in-distribution), concentrated in: invisible-Unicode (ASCII smuggling), base64-encoded payloads, and out-of-domain (non-coding) harm. An input-normalization defense closes the obfuscation gap. See the repo README for the full analysis.
python
import torchfrom transformers import AutoModelForSequenceClassification, AutoTokenizerfrom peft import PeftModel base = "Qwen/Qwen3-0.6B-Base"m = AutoModelForSequenceClassification.from_pretrained(base, num_labels=2, dtype=torch.float32)m = PeftModel.from_pretrained(m, "Aditi20/tool-guardrail-qwen3-0.6b-lora").merge_and_unload()tok = AutoTokenizer.from_pretrained("Aditi20/tool-guardrail-qwen3-0.6b-lora")Trained for the Repello AI research-engineer assignment. Not production-hardened.