GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Rakancorle1
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Rakancorle1
Model Tree
Qwen/Qwen3-4B-Instruct-2507
Input Modalities
Output Modalities
Supported Functionality
violation / no-violation| Model | Type | Params | Accuracy | F1 | Latency (ms/example) |
|---|---|---|---|---|---|
| PolicyGuard-4B (ours) | Guardrail | 4B | 0.9014 | 0.8759 | 22.5 |
| Llama-Guard-3-8B | Guardrail | 8B | 0.4246 | 0.5952 | 164.8 |
| Qwen-3-8B | Foundation | 8B | 0.6408 | 0.6407 | 115.8 |
| Llama-3.3-70B | Foundation | 70B | 0.9054 | 0.8883 | 305.0 |
| Gemini-1.5-Pro | Frontier | – | 0.8713 | 0.8502 | 596.1 |
➡️ PolicyGuard-4B achieves near-frontier accuracy at <1/10 inference cost, demonstrating that efficient, small-scale guardrails are feasible:contentReference[oaicite:1]{index=1}.
python
from transformers import AutoTokenizer, AutoModelForCausalLMimport torch, re model_id = "Rakancorle1/PolicyGuard-4B"tokenizer = AutoTokenizer.from_pretrained(model_id)model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto") def classify(policy, actions): if isinstance(actions, list): actions = "\n".join(actions) prompt = f"Policy: {policy}\n\nTrajectory Actions:\n{actions}\n\nOutput violation or no_violation." inputs = tokenizer(prompt, return_tensors="pt").to(model.device) outputs = model.generate(**inputs, max_new_tokens=4) text = tokenizer.decode(outputs[0][inputs['input_ids'].shape[1]:], skip_special_tokens=True).lower() return "no_violation" if "no_violation" in text else "violation" # Examplepolicy = "Do not submit a form without filling mandatory fields."actions = ["Open form page", "Click submit without input"]print(classify(policy, actions))# -> 'violation'from transformers import AutoTokenizer, AutoModelForCausalLMimport torch, re model_id = "Rakancorle1/PolicyGuard-4B"tokenizer = AutoTokenizer.from_pretrained(model_id)model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto") def classify(policy, actions): if isinstance(actions, list): actions = "\n".join(actions) prompt = f"Policy: {policy}\n\nTrajectory Actions:\n{actions}\n\nOutput violation or no_violation." inputs = tokenizer(prompt, return_tensors="pt").to(model.device) outputs = model.generate(**inputs, max_new_tokens=4) text = tokenizer.decode(outputs[0][inputs['input_ids'].shape[1]:], skip_special_tokens=True).lower() return "no_violation" if "no_violation" in text else "violation" # Examplepolicy = "Do not submit a form without filling mandatory fields."actions = ["Open form page", "Click submit without input"]print(classify(policy, actions))# -> 'violation'