GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Alhasan
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Alhasan
Model Tree
Qwen/Qwen3-1.7B
Input Modalities
Output Modalities
Supported Functionality
The model was trained with this exact system prompt and no examples. Using a different prompt degrades it badly:
text
You are the language-understanding module of a voice assistant for Arabic speakers. Reply with JSON only: {"intent": "<intent>", "slots": {"<slot_type>": "<value>"}}.The user turn is the raw command, nothing else. Valid intent and slot names are in
configs/schema.json in the GitHub repo.
python
from peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizer SYSTEM = ('You are the language-understanding module of a voice assistant for Arabic speakers. ' 'Reply with JSON only: {"intent": "<intent>", "slots": {"<slot_type>": "<value>"}}.') tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-1.7B")model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-1.7B", torch_dtype="bfloat16", device_map="auto")model = PeftModel.from_pretrained(model, "Alhasan/qwen3-1.7b-egyptian-arabic-lora") messages = [{"role": "system", "content": SYSTEM}, {"role": "user", "content": "الجو عامل ايه في اسكندرية النهارده"}]prompt = tok.apply_chat_template(messages, add_generation_prompt=True, tokenize=False, enable_thinking=False)out = model.generate(**tok(prompt, return_tensors="pt", add_special_tokens=False).to(model.device), max_new_tokens=128, do_sample=False)print(tok.decode(out[0][-128:], skip_special_tokens=True))Exact match = intent and every slot correct, on 200 native-written Egyptian commands:
| Model | Egyptian test | MASSIVE test (Saudi/MSA) |
|---|---|---|
| This adapter (Qwen3-1.7B + LoRA) | 0.590 | 0.583 |
| Qwen3-0.6B + LoRA | 0.510 | 0.539 |
| Claude Sonnet 5, 5-shot | 0.435 | 0.450 |
| Qwen3-0.6B untrained, 5-shot | 0.040 | 0.030 |
Fine-tuning beats the frontier model prompted with five examples by +15.5 points (95% CI [+8.5, +22.5], paired bootstrap). JSON validity is 100% with plain greedy decoding.
Converted with mlx_lm.convert; measured on an M4, 50 commands, one at a time:
| Build | Egyptian test | MASSIVE test | Disk | p50 latency |
|---|---|---|---|---|
| GPU (transformers, bf16) | 0.590 | 0.583 | — | — |
| MLX fp16 | 0.595 | 0.577 | 3.2 GB | 881 ms |
| MLX 4-bit | 0.555 | 0.577 | 948 MB | 345 ms |
The fp16 MLX build reproduces the GPU run to within one or two items, so the port is faithful. 4-bit has no net cost on MSA and costs about 4 points on Egyptian (−4.0 pts, 95% CI [−8.0, 0.0], p = 0.10, not significant at n=200). Quantization changes ~22% of outputs on both test sets. On MASSIVE the changes cancel out (11 items fixed, 11 broken). On Egyptian they lean negative (5 fixed, 13 broken). The lost items are mostly intent errors, spread evenly across dialect-marked and unmarked sentences. Use 4-bit for MSA freely; validate on a larger dialect set before shipping it for Egyptian.
LoRA rank 16, alpha 32, dropout 0.05, all attention and MLP projections; 1,200 optimizer steps at
batch 16, lr 2e-4 cosine, max length 256, loss on the assistant JSON only (completion-only).
Trained with TRL SFTTrainer on a single GPU (~1 h).
Full methodology, significance testing and figures: GitHub repo
from peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizer SYSTEM = ('You are the language-understanding module of a voice assistant for Arabic speakers. ' 'Reply with JSON only: {"intent": "<intent>", "slots": {"<slot_type>": "<value>"}}.') tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-1.7B")model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-1.7B", torch_dtype="bfloat16", device_map="auto")model = PeftModel.from_pretrained(model, "Alhasan/qwen3-1.7b-egyptian-arabic-lora") messages = [{"role": "system", "content": SYSTEM}, {"role": "user", "content": "الجو عامل ايه في اسكندرية النهارده"}]prompt = tok.apply_chat_template(messages, add_generation_prompt=True, tokenize=False, enable_thinking=False)out = model.generate(**tok(prompt, return_tensors="pt", add_special_tokens=False).to(model.device), max_new_tokens=128, do_sample=False)print(tok.decode(out[0][-128:], skip_special_tokens=True))