GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Tuan050901
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
Tuan050901
Model Tree
Qwen/Qwen2.5-1.5B-Instruct
Input Modalities
Output Modalities
Supported Functionality
Not a general chat model. Not a hate-speech classifier for protected attributes beyond what appears in the training mix.
python
import torchfrom peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizer base_id = "Qwen/Qwen2.5-1.5B-Instruct"adapter_id = "Tuan050901/vulgar-en-qlora" tokenizer = AutoTokenizer.from_pretrained(adapter_id, trust_remote_code=True)base = AutoModelForCausalLM.from_pretrained( base_id, torch_dtype=torch.bfloat16, device_map="auto", trust_remote_code=True,)model = PeftModel.from_pretrained(base, adapter_id)model.eval() system = ( "You are an English vulgar-language detector. " "Read the user text and return JSON only with keys " "is_vulgar, severity, and vulgar_terms. " "is_vulgar is true when the text uses English profanity, slurs, or obscene slang. " "severity is one of none, mild, strong, explicit. " "vulgar_terms is a list of the vulgar words or phrases found, in lowercase, " "normalized without leetspeak or masking characters. " "If the text is clean, set is_vulgar to false, severity to none, and vulgar_terms to []. " "Do not add commentary outside JSON.")text = "What the fuck is wrong with this compiler?"messages = [ {"role": "system", "content": system}, {"role": "user", "content": f"Text: {text}"},]prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)inputs = tokenizer(prompt, return_tensors="pt").to(model.device)with torch.inference_mode(): out = model.generate(**inputs, max_new_tokens=128, do_sample=False)print(tokenizer.decode(out[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))| Item | Value |
|---|---|
| Base | Qwen/Qwen2.5-1.5B-Instruct |
| Method | QLoRA 4-bit NF4 + LoRA r=16, alpha=32 |
| Targets | q/k/v/o/gate/up/down proj |
| Epochs | 3 |
| Hardware | 1x RTX 4090 24GB |
| Train time | ~1h 51m |
| Train samples | ~29,992 |
| Final train loss | 0.626 |
offensive, clean subset)Compared against the same base model without the adapter, same prompt/schema:
| Metric | Base | Fine-tuned | Delta |
|---|---|---|---|
| is_vulgar accuracy | 0.640 | 0.825 | +0.185 |
| is_vulgar F1 | 0.744 | 0.844 | +0.100 |
| severity macro-F1 | 0.170 | 0.741 | +0.571 |
| vulgar_terms F1 | 0.288 | 0.593 | +0.305 |
Full numbers: eval_compare.json in this repo.
| JSON parse failures | 3 | 0 | -3 |