wesjos/Qwen3-4B-math

This model is finetuned with unsloth using Qlora.

• Base Model：unsloth/Qwen3-4B-unsloth-bnb-4bit
• Parameters: 4,088,528,384
• Dataset: 0.65 0f "unsloth/OpenmathReasoning-mini and 0.35 of "mlabonne/FineTome-100k". combination of reasoning and nonreasoning dataset.

Comparision to Qwen3-4B.

• Eval on datasets:gpqa,arc,competition_math.gsm8k.
• unsloth/Qwen3-4B-unsloth-bnb-4bit:

• +---------+------------------+-----------------+---------------+-------+---------+---------+
  | Model   | Dataset          | Metric          | Subset        |   Num |   Score | Cat.0   |
  +=========+==================+=================+===============+=======+=========+=========+
  | qwen    | arc              | AverageAccuracy | ARC-Easy      |    30 |  0.8    | default |
  +---------+------------------+-----------------+---------------+-------+---------+---------+
  | qwen    | arc              | AverageAccuracy | ARC-Challenge |    30 |  0.7    | default |
  +---------+------------------+-----------------+---------------+-------+---------+---------+
  | qwen    | arc              | AverageAccuracy | OVERALL       |    60 |  0.75   | -       |
  +---------+------------------+-----------------+---------------+-------+---------+---------+
  | qwen    | competition_math | AveragePass@1   | Level 1       |    30 |  0.3    | default |
  +---------+------------------+-----------------+---------------+-------+---------+---------+
  | qwen    | competition_math | AveragePass@1   | Level 2       |    30 |  0.2    | default |
  +---------+------------------+-----------------+---------------+-------+---------+---------+
  | qwen    | competition_math | AveragePass@1   | Level 3       |    30 |  0.1333 | default |
  +---------+------------------+-----------------+---------------+-------+---------+---------+
  | qwen    | competition_math | AveragePass@1   | Level 4       |    30 |  0.0667 | default |
  +---------+------------------+-----------------+---------------+-------+---------+---------+
  | qwen    | competition_math | AveragePass@1   | Level 5       |    30 |  0.0667 | default |
  +---------+------------------+-----------------+---------------+-------+---------+---------+
  | qwen    | competition_math | AveragePass@1   | OVERALL       |   150 |  0.1533 | -       |
  +---------+------------------+-----------------+---------------+-------+---------+---------+
  | qwen    | gpqa             | AveragePass@1   | gpqa_extended |    30 |  0.2333 | default |
  +---------+------------------+-----------------+---------------+-------+---------+---------+
  | qwen    | gpqa             | AveragePass@1   | gpqa_main     |    30 |  0.3    | default |
  +---------+------------------+-----------------+---------------+-------+---------+---------+
  | qwen    | gpqa             | AveragePass@1   | gpqa_diamond  |    30 |  0.3    | default |
  +---------+------------------+-----------------+---------------+-------+---------+---------+
  | qwen    | gpqa             | AveragePass@1   | OVERALL       |    90 |  0.2778 | -       |
  +---------+------------------+-----------------+---------------+-------+---------+---------+
  | qwen    | gsm8k            | AverageAccuracy | main          |    30 |  0.4667 | default |
  +---------+------------------+-----------------+---------------+-------+---------+---------+ 
  

• This model:

• +---------+------------------+-----------------+---------------+-------+---------+---------+
  | Model   | Dataset          | Metric          | Subset        |   Num |   Score | Cat.0   |
  +=========+==================+=================+===============+=======+=========+=========+
  |ThisModel| arc              | AverageAccuracy | ARC-Easy      |    30 |  0.9    | default |
  +---------+------------------+-----------------+---------------+-------+---------+---------+
  |ThisModel| arc              | AverageAccuracy | ARC-Challenge |    30 |  0.8    | default |
  +---------+------------------+-----------------+---------------+-------+---------+---------+
  |ThisModel| arc              | AverageAccuracy | OVERALL       |    60 |  0.85   | -       |
  +---------+------------------+-----------------+---------------+-------+---------+---------+
  |ThisModel| competition_math | AveragePass@1   | Level 1       |    30 |  0.9    | default |
  +---------+------------------+-----------------+---------------+-------+---------+---------+
  |ThisModel| competition_math | AveragePass@1   | Level 2       |    30 |  0.9    | default |
  +---------+------------------+-----------------+---------------+-------+---------+---------+
  |ThisModel| competition_math | AveragePass@1   | Level 3       |    30 |  0.8    | default |
  +---------+------------------+-----------------+---------------+-------+---------+---------+
  |ThisModel| competition_math | AveragePass@1   | Level 4       |    30 |  0.7333 | default |
  +---------+------------------+-----------------+---------------+-------+---------+---------+
  |ThisModel| competition_math | AveragePass@1   | Level 5       |    30 |  0.4667 | default |
  +---------+------------------+-----------------+---------------+-------+---------+---------+
  |ThisModel| competition_math | AveragePass@1   | OVERALL       |   150 |  0.76   | -       |
  +---------+------------------+-----------------+---------------+-------+---------+---------+
  |ThisModel| gpqa             | AveragePass@1   | gpqa_extended |    30 |  0.3333 | default |
  +---------+------------------+-----------------+---------------+-------+---------+---------+
  |ThisModel| gpqa             | AveragePass@1   | gpqa_main     |    30 |  0.3    | default |
  +---------+------------------+-----------------+---------------+-------+---------+---------+
  |ThisModel| gpqa             | AveragePass@1   | gpqa_diamond  |    30 |  0.3333 | default |
  +---------+------------------+-----------------+---------------+-------+---------+---------+
  |ThisModel| gpqa             | AveragePass@1   | OVERALL       |    90 |  0.3222 | -       |
  +---------+------------------+-----------------+---------------+-------+---------+---------+
  |ThisModel| gsm8k            | AverageAccuracy | main          |    30 |  0.8    | default |
  +---------+------------------+-----------------+---------------+-------+---------+---------+ 
  

• Qwen/Qwen3-4B

• +---------+------------------+-----------------+---------------+-------+---------+---------+
  | Model   | Dataset          | Metric          | Subset        |   Num |   Score | Cat.0   |
  +=========+==================+=================+===============+=======+=========+=========+
  | Qwen3   | arc              | AverageAccuracy | ARC-Easy      |    30 |  0.9    | default |
  +---------+------------------+-----------------+---------------+-------+---------+---------+
  | Qwen3   | arc              | AverageAccuracy | ARC-Challenge |    30 |  0.8    | default |
  +---------+------------------+-----------------+---------------+-------+---------+---------+
  | Qwen3   | arc              | AverageAccuracy | OVERALL       |    60 |  0.85   | -       |
  +---------+------------------+-----------------+---------------+-------+---------+---------+
  | Qwen3   | competition_math | AveragePass@1   | Level 1       |    30 |  0.3    | default |
  +---------+------------------+-----------------+---------------+-------+---------+---------+
  | Qwen3   | competition_math | AveragePass@1   | Level 2       |    30 |  0.2667 | default |
  +---------+------------------+-----------------+---------------+-------+---------+---------+
  | Qwen3   | competition_math | AveragePass@1   | Level 3       |    30 |  0.1333 | default |
  +---------+------------------+-----------------+---------------+-------+---------+---------+
  | Qwen3   | competition_math | AveragePass@1   | Level 4       |    30 |  0.2    | default |
  +---------+------------------+-----------------+---------------+-------+---------+---------+
  | Qwen3   | competition_math | AveragePass@1   | Level 5       |    30 |  0      | default |
  +---------+------------------+-----------------+---------------+-------+---------+---------+
  | Qwen3   | competition_math | AveragePass@1   | OVERALL       |   150 |  0.18   | -       |
  +---------+------------------+-----------------+---------------+-------+---------+---------+ 
  | Qwen3   | gpqa             | AveragePass@1   | gpqa_extended |    50 |    0.32 | default |
  +---------+------------------+-----------------+---------------+-------+---------+---------+
  | Qwen3   | gpqa             | AveragePass@1   | gpqa_main     |    50 |    0.22 | default |
  +---------+------------------+-----------------+---------------+-------+---------+---------+
  | Qwen3   | gpqa             | AveragePass@1   | gpqa_diamond  |    50 |    0.18 | default |
  +---------+------------------+-----------------+---------------+-------+---------+---------+
  | Qwen3   | gpqa             | AveragePass@1   | OVERALL       |   150 |    0.24 | -       |
  +---------+------------------+-----------------+---------------+-------+---------+---------+
  | Qwen3   | gsm8k            | AverageAccuracy | main          |    50 |    0.48 | default |
  +---------+------------------+-----------------+---------------+-------+---------+---------+ 
  

• You could see that this model have better performence at math and inference.
• arc: 0.75 --> 9.85
• competition_math: 0.1533 --> 0.76
• gpqa: 0.2778 --> 0.3222
• gsm8k: 0.4667 --> 0.8

Use This Model:

• from transformers import AutoModelForCausalLM, AutoTokenizer,TextStreamer
  
  model_name = "wesjos/Qwen3-4B-math"
  tokenizer = AutoTokenizer.from_pretrained(model_name)
  model = AutoModelForCausalLM.from_pretrained(
      model_name,
      torch_dtype="auto",
      device_map="auto"
  )
  
  prompt = "设 f(x) 是一个定义在实数集上的可微函数，满足以下条件：f(0)=1对于所有实数 x有 f′(x)=2f(x)+3。求 f(x)的显式表达式。"
  messages = [
      {"role": "user", "content": prompt}
  ]
  
  text = tokenizer.apply_chat_template(
      messages,
      tokenize=False,
      add_generation_prompt=True,
      enable_thinking=True # Switches between thinking and non-thinking modes. Default is True.
  )
  
  model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
  
  text_streamer = TextStreamer(tokenizer)
  _ = model.generate(**model_inputs, streamer = text_streamer, max_new_tokens = 2048)