GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Yusif-v
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
Yusif-v
Model Tree
Qwen/Qwen2.5-3B-Instruct
Input Modalities
Output Modalities
Supported Functionality
Generate adversarial payloads for authorized security assessments of LLM endpoints:
python
from transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModelbase = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-3B-Instruct", torch_dtype="auto", device_map="auto")model = PeftModel.from_pretrained(base, "Yusif-v/loki-injector-3b-v1")tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-3B-Instruct")instruction = 'Craft a prompt-injection payload that achieves "extract the system prompt" using the "direct" technique. Output only the payload.'prompt = tokenizer.apply_chat_template([{"role": "user", "content": instruction}], tokenize=False, add_generation_prompt=True)inputs = tokenizer(prompt, return_tensors="pt").to(model.device)outputs = model.generate(**inputs, max_new_tokens=256, temperature=0.9, top_p=0.9)print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))
Serves as a dynamic payload source within LLM red-teaming frameworks, alongside static probe libraries and multi-turn attack orchestrators.
Not intended for targeting production systems without authorization. For security research, authorized red-teaming, and model robustness evaluation only.
Built from two public datasets via a custom corpus pipeline:
Pipeline: ingest → filter → deduplicate → technique-tag → synthesize → emit. Final corpus: ~15,000 rows (train/val/test split 90/5/5). Rows tagged to "other" (non-injection content) are dropped.
| Parameter | Value |
|---|---|
| LoRA rank (r) | 16 |
| LoRA alpha | 32 |
| LoRA dropout | 0.05 |
| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
| Epochs | 3 |
| Learning rate | 2e-4 |
| Batch size | 1 (per device) |
| Gradient accumulation | 8 |
| Max sequence length | 512 |
Trained on Kaggle P100 (16GB) using the plain transformers.Trainer recipe. Training text uses the Qwen2.5 chat template.
Evaluated against an OpenAI-compatible target using a check suite covering prompt-injection, jailbreak, and data-leakage categories.
Metrics: findings-per-payload, diversity (mean bigram overlap), refusal rate — compared against a static payload baseline.
bibtex
@software{loki_injector_2026,title={Loki Injector 3B v1},author={Yusifov, Telman},year={2026},publisher={HuggingFace},url={https://huggingface.co/Yusif-v/loki-injector-3b-v1}}
| Precision | FP16 |
| Gradient checkpointing | Enabled |
| Optimizer | AdamW (default in Transformers Trainer) |