GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
itsjesusmacias
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
itsjesusmacias
Model Tree
Qwen/Qwen3-0.6B
Input Modalities
Output Modalities
Supported Functionality
Given a free-text incident message from a student (in Spanish), the model produces:
json
{ "category": "billing | account | technical | other", "priority": "low | medium | high", "response": "<brief support response in Spanish>"}| Parameter | Value |
|---|---|
r | 8 |
lora_alpha | 16 |
lora_dropout | 0.05 |
target_modules | q_proj, v_proj |
bias | none |
task_type | CAUSAL_LM |
| Parameter | Value |
|---|---|
| Epochs | 3 |
| Learning rate | 2e-4 |
| Batch size (effective) | 8 (1 x grad_accum 8) |
| Optimizer | AdamW (adamw_torch) |
| Warmup ratio | 0.1 |
| Max grad norm | 1.0 |
| Max seq length | 600 tokens |
| Seed | 7 |
| Condition | Valid JSON | Valid contract | Category acc. | Priority acc. |
|---|---|---|---|---|
| Base (no fine-tuning) | 96.7% | 96.7% | 36.7% | 36.7% |
| Few-shot (4 examples) | 100.0% | 100.0% | 63.3% | 56.7% |
| LoRA (this adapter) | 100.0% | 100.0% | 73.3% | 63.3% |
LoRA improves category accuracy by +36.6 pp and priority accuracy by +26.6 pp over the base model.
python
import torchfrom peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizer model_id = "Qwen/Qwen3-0.6B"adapter_id = "itsjesusmacias/qwen3-0.6b-lora-incidents" tokenizer = AutoTokenizer.from_pretrained(model_id)base_model = AutoModelForCausalLM.from_pretrained( model_id, dtype=torch.float32, device_map="auto",)model = PeftModel.from_pretrained(base_model, adapter_id)model.eval() SYSTEM_PROMPT = ( "Eres un asistente de soporte de una plataforma educativa online. " "Recibes el texto de una incidencia y debes clasificarla en JSON con exactamente estos campos: " "category (billing, account, technical, other), " "priority (low, medium, high), " "response (respuesta breve al usuario en espanol). " "Responde SOLO con el JSON, sin texto adicional.") message = "No puedo acceder a mi cuenta y tengo un examen en una hora."messages = [ {"role": "system", "content": SYSTEM_PROMPT}, {"role": "user", "content": message},] prompt = tokenizer.apply_chat_template( messages, tokenize=False, add_generation_prompt=True, enable_thinking=False,)inputs = tokenizer(prompt, return_tensors="pt").to(model.device) with torch.inference_mode(): output = model.generate( **inputs, max_new_tokens=256, do_sample=True, temperature=0.7, top_p=0.8, top_k=20, ) new_tokens = output[0, inputs["input_ids"].shape[1]:]print(tokenizer.decode(new_tokens, skip_special_tokens=True))# {"category": "account", "priority": "high", "response": "..."}| Parameter | Value |
|---|---|
max_new_tokens | 256 |
temperature | 0.7 |
top_p | 0.8 |
top_k | 20 |
enable_thinking | False |
enable_thinking=False in the Qwen3 chat template. Always pass enable_thinking=False when applying the template, otherwise the model will enter thinking mode and the output format may differ.Apache 2.0 (inherited from base model Qwen/Qwen3-0.6B)
{ "category": "billing | account | technical | other", "priority": "low | medium | high", "response": "<brief support response in Spanish>"}import torchfrom peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizer model_id = "Qwen/Qwen3-0.6B"adapter_id = "itsjesusmacias/qwen3-0.6b-lora-incidents" tokenizer = AutoTokenizer.from_pretrained(model_id)base_model = AutoModelForCausalLM.from_pretrained( model_id, dtype=torch.float32, device_map="auto",)model = PeftModel.from_pretrained(base_model, adapter_id)model.eval() SYSTEM_PROMPT = ( "Eres un asistente de soporte de una plataforma educativa online. " "Recibes el texto de una incidencia y debes clasificarla en JSON con exactamente estos campos: " "category (billing, account, technical, other), " "priority (low, medium, high), " "response (respuesta breve al usuario en espanol). " "Responde SOLO con el JSON, sin texto adicional.") message = "No puedo acceder a mi cuenta y tengo un examen en una hora."messages = [ {"role": "system", "content": SYSTEM_PROMPT}, {"role": "user", "content": message},] prompt = tokenizer.apply_chat_template( messages, tokenize=False, add_generation_prompt=True, enable_thinking=False,)inputs = tokenizer(prompt, return_tensors="pt").to(model.device) with torch.inference_mode(): output = model.generate( **inputs, max_new_tokens=256, do_sample=True, temperature=0.7, top_p=0.8, top_k=20, ) new_tokens = output[0, inputs["input_ids"].shape[1]:]print(tokenizer.decode(new_tokens, skip_special_tokens=True))# {"category": "account", "priority": "high", "response": "..."}