---
base_model:
- unsloth/Qwen3-8B
tags:
- text-generation-inference
- transformers
- unsloth
- qwen3
- urdu
- urdu-reasoning
license: apache-2.0
language:
- ur
widget:
- messages:
  - role: user
    content: 2 + 3 کتنا ہوتا ہے؟
- messages:
  - role: user
    content: اگر آپ کے پاس 10 سیب ہیں اور 3 کھا لیتے ہیں، تو کتنے بچیں گے؟
- messages:
  - role: user
    content: >-
      ایک باپ کی عمر اپنے بیٹے کی عمر سے تین گنا ہے۔ 10 سال بعد باپ کی عمر 50
      سال ہوگی۔ بیٹے کی موجودہ عمر کیا ہے؟
- messages:
  - role: user
    content: >-
      اگر ایک گاڑی 60 کلومیٹر فی گھنٹہ کی رفتار سے 2 گھنٹے چلے، تو کتنی دوری طے
      کرے گی؟
---




## Quick start

```python
from unsloth import FastLanguageModel
import torch

max_seq_length = 2048  # Choose any! We auto support RoPE Scaling internally!
dtype = (
    None  # None for auto detection. Float16 for Tesla T4, V100, Bfloat16 for Ampere+
)
load_in_4bit = False  # Use 4bit quantization to reduce memory usage. Can be False.
load_in_8bit = False  # Use 8bit quantization to reduce memory usage. Can be False.

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="azherali/Riazi-8B-Instruct",  # Choose ANY
    max_seq_length=max_seq_length,
    dtype=dtype,
    load_in_4bit=load_in_4bit,
    load_in_8bit=load_in_8bit,
    # token = "YOUR_HF_TOKEN", # HF Token for gated models
)
FastLanguageModel.for_inference(model)  # Enable native 2x faster inference

reasoning_start = "<reasoning>"
reasoning_end   = "</reasoning>"
solution_start = "<SOLUTION>"
solution_end = "</SOLUTION>"

system_prompt = f"""
You are given a problem.

Think about the problem and provide your working out.
Place your reasoning between {reasoning_start} and {reasoning_end}.

Then, provide your final solution between
{solution_start} and {solution_end}.

Always answer in Urdu.
"""

message = [
     {"role": "system", "content": system_prompt},
    {
        "role": "user",
        "content": "پانچ بچوں نے 20 چاکلیٹس برابر بانٹیں۔ ہر بچے کو کتنی چاکلیٹس ملیں گی؟"
    }
]
text = tokenizer.apply_chat_template(
    message,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False,
    # Must add for generation
)

# Inference Using Tranformer Library
from transformers import TextStreamer

_ = model.generate(
    **tokenizer(text, return_tensors="pt").to("cuda"),
    temperature=0.6,
    top_p=0.95,
    top_k=20,  # For non thinking
    streamer=TextStreamer(tokenizer, skip_prompt=True),
)

# Inference Using VLLM(fast)
from vllm import SamplingParams
def generate_answer(problem):
    message = [
        {"role": "system", "content": system_prompt},
        {"role": "user",   "content": problem},
    ]

    text = tokenizer.apply_chat_template(
        message,
        add_generation_prompt = True, # Must add for generation
        tokenize = False,
        enable_thinking=False,
    )
    sampling_params = SamplingParams(
        max_tokens=2048,
        temperature=0.6,
        top_p=0.95,
        top_k=20,
    )
    output = model.fast_generate(
        text,
        sampling_params = sampling_params,
        lora_request =None,
    )[0].outputs[0].text
    return output

generate_answer("پانچ بچوں نے 20 چاکلیٹس برابر بانٹیں۔ ہر بچے کو کتنی چاکلیٹس ملیں گی؟")
```

## Training procedure

 


This model was trained with SFT.

### Framework versions

- TRL: 0.22.2
- Transformers: 4.56.2
- Pytorch: 2.12.0+rocm7.2
- Datasets: 4.3.0
- Tokenizers: 0.22.2