GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
abdullahashraf122
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
abdullahashraf122
Model Tree
Qwen/Qwen2.5-3B-Instruct
Input Modalities
Output Modalities
Supported Functionality
This model repository has been certified under formal empirical security evaluation standards with cryptographic attestation certificates sealed via keyed HMAC-SHA256 bound to the model weight digest (2817e14ac2...) and Git commit (7afc255278...).
👉 Read the Full Model Security Card: SECURITY_CARD.md
| Benchmark Standard | Target Metric | Certified Empirical Result | Standard Mandate | Determination |
|---|---|---|---|---|
| Meta Purple Llama CyberSecEval 3 | Attack Success Rate (ASR) | 0.0% (100% Contained) | ≤0.1% | PASSED |
| Meta CyberSecEval Containment | Guardrail Refusal Rate | 10.0% (10 Refused / 90 Compliant) | Targeted Guard | PASSED |
| Structured Output Lock | Pydantic v2 Schema Conformance | 100.0% | 100.0% | PASSED |
| Conformal Prediction Bound | Empirical False Positive Rate | 0.0000% | ≤0.1% at 95% Conf | PASSED |
| Digital Dubai AI Ethics | Demographic Disparity Ratio | 0.8523 (Four-Fifths Rule) | ≥0.80 | PASSED |
| Dubai DESC ISR v2.0 | Data Sovereignty & Residency | Certified Tier 1 & Tier 2 Scope | 100% On-Premise / Zero Offshore | PASSED |
hares-qwen2.5-3b is an instruction-tuned small language model (SLM) based on Qwen/Qwen2.5-3B-Instruct, specialized in detecting social engineering, impersonation scams, phishing lures, and financial extortions targeting residents of Dubai and the UAE across Arabic (Emirati Gulf, Modern Standard Arabic), Expat English, and Arabizi (Franco-Arabic).
The model was fine-tuned via QLoRA FP16 on 100,000 localized UAE cyber fraud cases (uae_smishing_100k.jsonl), merged into pure Float16 SafeTensors shards (4 shards, 434 tensor layers, 3,085,938,688 parameters, zero Python pickle code).
This model is intended to be used with:
The model produces structured JSON adhering to the TextTriageInferenceResult schema:
json
{ "is_impersonation": true, "confidence_score": 0.96, "suspected_entity": "Dubai Police", "tactic_category": "authority_impersonation", "reasoning_summary": "High Risk: Fraudulent traffic fine extortion impersonating Dubai Police with an unverified external payment link.", "escalation_recommended": true}{ "is_impersonation": true, "confidence_score": 0.96, "suspected_entity": "Dubai Police", "tactic_category": "authority_impersonation", "reasoning_summary": "High Risk: Fraudulent traffic fine extortion impersonating Dubai Police with an unverified external payment link.", "escalation_recommended": true}