GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
amwangfan
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
amwangfan
Model Tree
Qwen/Qwen2.5-0.5B-Instruct
Input Modalities
Output Modalities
Supported Functionality
generate() 方式，直接提取最后一个 Token 对应 SECRET (ID: 65310) 与 SAFE (ID: 83788) 的 Logits 分数进行相对概率比较。lm_head（rank=16, alpha=32）。lm_head（tie_word_embeddings = False），增强对分类目标 Token 的区分度。python
import torchimport torch.nn.functional as Ffrom transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModelbase_model_name = "Qwen/Qwen2.5-0.5B-Instruct"lora_model_name = "amwangfan/Qwen2.5-0.5B-Privacy-Gateway-v3-LoRA"tokenizer = AutoTokenizer.from_pretrained(base_model_name)base_model = AutoModelForCausalLM.from_pretrained(base_model_name, torch_dtype=torch.float16, device_map="auto")model = PeftModel.from_pretrained(base_model, lora_model_name)model.eval()TOKEN_ID_SECRET = 65310 # 'SECRET'TOKEN_ID_SAFE = 83788 # 'SAFE'def classify_credential(candidate_text: str, threshold: float = 0.5):prompt = ("判断以下词或短语是否为需要保密的隐私敏感信息""（包括密码、API密钥、访问令牌、私钥等），"f"仅回答 SECRET 或 SAFE：\n{candidate_text}\n答案：")inputs = tokenizer(prompt, return_tensors="pt").to(model.device)with torch.inference_mode():outputs = model(**inputs)next_logits = outputs.logits[0, -1, :].float()score_secret = next_logits[TOKEN_ID_SECRET]score_safe = next_logits[TOKEN_ID_SAFE]pair_probs = F.softmax(torch.stack([score_secret, score_safe]), dim=0)p_secret = pair_probs[0].item()return {"candidate": candidate_text,"is_sensitive": p_secret >= threshold,"p_secret": p_secret,"diff": (score_secret - score_safe).item()}print(classify_credential("sk-example-not-a-real-key-1234567890"))print(classify_credential("今天天气晴朗适合散步"))
完整网关（正则 Layer 0 + 本模型残差 + Vault 还原）见 GitHub，不要把整段对话丢进 0.5B。
<PRIV_{hash8}_{TYPE}_{N}>，出站 DFA 还原。Apache-2.0