Jupyter Agent Qwen3-4B Thinking

Jupyter Agent Qwen3-4B Thinking is a fine-tuned version of Qwen3-4B-Thinking-2507 specifically optimized for data science agentic tasks in Jupyter notebook environments. This model can execute Python code, analyze datasets, and provide step-by-step reasoning with intermediate computations to solve realistic data analysis problems.

• Model type: Causal Language Model (Thinking)
• Language(s): English, Python
• License: Apache 2.0
• Finetuned from: Qwen/Qwen3-4B-Thinking-2507

Key Features

• Jupyter-native agent that lives inside notebook environments
• Code execution with pandas, numpy, matplotlib, and other data science libraries
• Step-by-step reasoning with intermediate computations and thinking traces
• Dataset-grounded analysis trained on real Kaggle notebook workflows
• Tool calling for structured code execution and final answer generation

Performance

On the DABStep benchmark for data science tasks:

State-of-the-art performance for small models on realistic data analysis tasks.

Model Sources

• Repository: jupyter-agent
• Dataset: jupyter-agent-dataset
• Blog post: Jupyter Agents: training LLMs to reason with notebooks
• Demo: Jupyter Agent 2

Usage

Basic Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "jupyter-agent/jupyter-agent-qwen3-4b-thinking"

# Load model and tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

# Prepare input
prompt = "Analyze this sales dataset and find the top 3 performing products by revenue."
messages = [
    {"role": "user", "content": prompt}
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

# Generate response
generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=16384
)
output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist()

Decoding Thinking and Content

For thinking models, you can extract both the reasoning and final response:

try:
    # Find the end of thinking section (</think>)
    index = len(output_ids) - output_ids[::-1].index(151668)
except ValueError:
    index = 0

thinking_content = tokenizer.decode(output_ids[:index], skip_special_tokens=True).strip("\n")
content = tokenizer.decode(output_ids[index:], skip_special_tokens=True).strip("\n")

print("Thinking:", thinking_content)
print("Response:", content)

Agentic Usage with Tool Calling

The model works best with proper scaffolding for tool calling:

tools = [
    {
        "type": "function",
        "function": {
            "name": "execute_code",
            "description": "Execute Python code in a Jupyter environment",
            "parameters": {
                "type": "object",
                "properties": {
                    "code": {
                        "type": "string",
                        "description": "Python code to execute"
                    }
                },
                "required": ["code"]
            }
        }
    },
    {
        "type": "function", 
        "function": {
            "name": "final_answer",
            "description": "Provide the final answer to the question",
            "parameters": {
                "type": "object",
                "properties": {
                    "answer": {
                        "type": "string",
                        "description": "The final answer"
                    }
                },
                "required": ["answer"]
            }
        }
    }
]

# Include tools in the conversation
messages = [
    {
        "role": "system", 
        "content": "You are a data science assistant. Use the available tools to analyze data and provide insights."
    },
    {"role": "user", "content": prompt}
]

Training Details

Training Data

The model was fine-tuned on the Jupyter Agent Dataset, which contains:

• 51,389 synthetic notebooks (~0.2B tokens, total 1B tokens)
• Dataset-grounded QA pairs from real Kaggle notebooks
• Executable reasoning traces with intermediate computations
• High-quality educational content filtered and scored by LLMs

Training Procedure

• Base Model: Qwen3-4B-Thinking-2507
• Training Method: Full-parameter fine-tuning (not PEFT)
• Optimizer: AdamW with cosine learning rate scheduling
• Learning Rate: 5e-6
• Epochs: 5 (optimal based on ablation study)
• Context Length: 32,768 tokens
• Batch Size: Distributed across multiple GPUs
• Loss: Assistant-only loss (assistant_loss_only=True)
• Regularization: NEFTune noise (α=7) for full-parameter training

Training Infrastructure

• Framework: TRL with Transformers
• Distributed Training: DeepSpeed ZeRO-2 across multiple nodes
• Hardware: Multi-GPU setup with SLURM orchestration

Evaluation

Benchmark: DABStep

The model was evaluated on DABStep, a benchmark for data science agents with realistic tasks involving:

• Dataset analysis with pandas and numpy
• Visualization with matplotlib/seaborn
• Statistical analysis and business insights
• Multi-step reasoning with intermediate computations

The model achieves 26.8% improvement over the base model and 11.1% improvement over scaffolding alone.

We can also see, that the hard score can increase too even though our dataset is focused on easier questions.

Limitations and Bias

Technical Limitations

• Context window: Limited to 32K tokens, may struggle with very large notebooks
• Tool calling format: Requires specific scaffolding for optimal performance
• Dataset domains: Primarily trained on Kaggle-style data science tasks
• Code execution: Requires proper sandboxing for safe execution

Potential Biases

• Domain bias: Trained primarily on Kaggle notebooks, may not generalize to all data science workflows
• Language bias: Optimized for English and Python, limited multilingual support
• Task bias: Focused on structured data analysis, may underperform on unstructured data tasks

Recommendations

• Use in sandboxed environments like E2B for safe code execution
• Validate outputs before using in production systems
• Review generated code for security and correctness
• Consider domain adaptation for specialized use cases

Ethical Considerations

• Code Safety: Always execute generated code in secure, isolated environments
• Data Privacy: Be cautious when analyzing sensitive datasets
• Verification: Validate all analytical conclusions and insights
• Attribution: Acknowledge model assistance in data analysis workflows

Citation

@misc{jupyteragentqwen3thinking,
  title={Jupyter Agent Qwen3-4B Thinking},
  author={Baptiste Colle and Hanna Yukhymenko and Leandro von Werra},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/jupyter-agent/jupyter-agent-qwen3-4b-thinking}
}

Related Work

• Dataset: jupyter-agent-dataset
• Non-thinking version: jupyter-agent-qwen3-4b-instruct
• Base model: Qwen3-4B-Thinking-2507
• Benchmark: DABStep

For more details, see our blog post and GitHub repository.