gabriellarson/jupyter-agent-qwen3-4b-thinking-gguf overview

{
  "metadata": {},
  "card_data": {
    "license": "apache-2.0",
    "library_name": "transformers",
    "tags": [
      "code",
      "jupyter",
      "agent",
      "data-science",
      "qwen",
      "thinking"
    ],
    "base_model": [
      "jupyter-agent/jupyter-agent-qwen3-4b-thinking"
    ],
    "datasets": [
      "jupyter-agent/jupyter-agent-dataset"
    ],
    "language": [
      "en",
      "code"
    ],
    "pipeline_tag": "text-generation",
    "frontmatter": {
      "license": "apache-2.0",
      "library_name": "transformers",
      "tags": [
        "code",
        "jupyter",
        "agent",
        "data-science",
        "qwen",
        "thinking"
      ],
      "base_model": [
        "jupyter-agent/jupyter-agent-qwen3-4b-thinking"
      ],
      "datasets": [
        "jupyter-agent/jupyter-agent-dataset"
      ],
      "language": [
        "en",
        "code"
      ],
      "pipeline_tag": "text-generation"
    },
    "hero_image_url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/jupyter-agent-2/training_dabstep_easy.png",
    "summary": "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: `python try: # Find the end of thinking section () 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: `python 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} ] ``",
    "quick_links": [],
    "benchmark_table_html": "",
    "readme_markdown": "---\nlicense: apache-2.0\nlibrary_name: transformers\ntags:\n- code\n- jupyter\n- agent\n- data-science\n- qwen\n- thinking\nbase_model:\n- jupyter-agent/jupyter-agent-qwen3-4b-thinking\ndatasets:\n- jupyter-agent/jupyter-agent-dataset\nlanguage:\n- en\n- code\npipeline_tag: text-generation\n---\n# Jupyter Agent Qwen3-4B Thinking\n\n![image/png](https://cdn-uploads.huggingface.co/production/uploads/650ed7adf141bc34f91a12ae/ZyF9foqe5SLECwkq0dOpT.png)\n\n**Jupyter Agent Qwen3-4B Thinking** is a fine-tuned version of [Qwen3-4B-Thinking-2507](https://huggingface.co/Qwen/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.\n\n- **Model type:** Causal Language Model (Thinking)\n- **Language(s):** English, Python\n- **License:** Apache 2.0\n- **Finetuned from:** [Qwen/Qwen3-4B-Thinking-2507](https://huggingface.co/Qwen/Qwen3-4B-Thinking-2507)\n\n## Key Features\n\n- **Jupyter-native agent** that lives inside notebook environments\n- **Code execution** with pandas, numpy, matplotlib, and other data science libraries\n- **Step-by-step reasoning** with intermediate computations and thinking traces\n- **Dataset-grounded analysis** trained on real Kaggle notebook workflows\n- **Tool calling** for structured code execution and final answer generation\n\n## Performance\n\nOn the [DABStep benchmark](https://huggingface.co/spaces/adyen/DABstep) for data science tasks:\n\n| Model | Easy Tasks | Hard Tasks |\n|-------|------------|------------|\n| Qwen3-4B-Thinking-2507 (Base) | 44.0% | 2.1% |\n| **Jupyter Agent Qwen3-4B Thinking** | **70.8%** | **3.4%** |\n\n**State-of-the-art performance** for small models on realistic data analysis tasks.\n\n## Model Sources\n\n- **Repository:** [jupyter-agent](https://github.com/huggingface/jupyter-agent)\n- **Dataset:** [jupyter-agent-dataset](https://huggingface.co/datasets/jupyter-agent/jupyter-agent-dataset)\n- **Blog post:** [Jupyter Agents: training LLMs to reason with notebooks](https://huggingface.co/blog/jupyter-agent-2)\n- **Demo:** [Jupyter Agent 2](https://huggingface.co/spaces/lvwerra/jupyter-agent-2)\n\n## Usage\n\n### Basic Usage\n\n```python\nfrom transformers import AutoModelForCausalLM, AutoTokenizer\nmodel_name = \"jupyter-agent/jupyter-agent-qwen3-4b-thinking\"\n# Load model and tokenizer\ntokenizer = AutoTokenizer.from_pretrained(model_name)\nmodel = AutoModelForCausalLM.from_pretrained(\n    model_name,\n    torch_dtype=\"auto\",\n    device_map=\"auto\"\n)\n# Prepare input\nprompt = \"Analyze this sales dataset and find the top 3 performing products by revenue.\"\nmessages = [\n    {\"role\": \"user\", \"content\": prompt}\n]\ntext = tokenizer.apply_chat_template(\n    messages,\n    tokenize=False,\n    add_generation_prompt=True\n)\nmodel_inputs = tokenizer([text], return_tensors=\"pt\").to(model.device)\n# Generate response\ngenerated_ids = model.generate(\n    **model_inputs,\n    max_new_tokens=16384\n)\noutput_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist()\n```\n\n### Decoding Thinking and Content\n\nFor thinking models, you can extract both the reasoning and final response:\n\n```python\ntry:\n    # Find the end of thinking section (</think>)\n    index = len(output_ids) - output_ids[::-1].index(151668)\nexcept ValueError:\n    index = 0\nthinking_content = tokenizer.decode(output_ids[:index], skip_special_tokens=True).strip(\"\\n\")\ncontent = tokenizer.decode(output_ids[index:], skip_special_tokens=True).strip(\"\\n\")\nprint(\"Thinking:\", thinking_content)\nprint(\"Response:\", content)\n```\n\n### Agentic Usage with Tool Calling\n\nThe model works best with proper scaffolding for tool calling:\n\n```python\ntools = [\n    {\n        \"type\": \"function\",\n        \"function\": {\n            \"name\": \"execute_code\",\n            \"description\": \"Execute Python code in a Jupyter environment\",\n            \"parameters\": {\n                \"type\": \"object\",\n                \"properties\": {\n                    \"code\": {\n                        \"type\": \"string\",\n                        \"description\": \"Python code to execute\"\n                    }\n                },\n                \"required\": [\"code\"]\n            }\n        }\n    },\n    {\n        \"type\": \"function\", \n        \"function\": {\n            \"name\": \"final_answer\",\n            \"description\": \"Provide the final answer to the question\",\n            \"parameters\": {\n                \"type\": \"object\",\n                \"properties\": {\n                    \"answer\": {\n                        \"type\": \"string\",\n                        \"description\": \"The final answer\"\n                    }\n                },\n                \"required\": [\"answer\"]\n            }\n        }\n    }\n]\n# Include tools in the conversation\nmessages = [\n    {\n        \"role\": \"system\", \n        \"content\": \"You are a data science assistant. Use the available tools to analyze data and provide insights.\"\n    },\n    {\"role\": \"user\", \"content\": prompt}\n]\n```\n\n## Training Details\n\n### Training Data\n\nThe model was fine-tuned on the [Jupyter Agent Dataset](https://huggingface.co/datasets/jupyter-agent/jupyter-agent-dataset), which contains:\n\n- **51,389 synthetic notebooks** (~0.2B tokens, total 1B tokens)\n- **Dataset-grounded QA pairs** from real Kaggle notebooks\n- **Executable reasoning traces** with intermediate computations\n- **High-quality educational content** filtered and scored by LLMs\n\n### Training Procedure\n\n- **Base Model:** Qwen3-4B-Thinking-2507\n- **Training Method:** Full-parameter fine-tuning (not PEFT)\n- **Optimizer:** AdamW with cosine learning rate scheduling\n- **Learning Rate:** 5e-6\n- **Epochs:** 5 (optimal based on ablation study)\n- **Context Length:** 32,768 tokens\n- **Batch Size:** Distributed across multiple GPUs\n- **Loss:** Assistant-only loss (`assistant_loss_only=True`)\n- **Regularization:** NEFTune noise (α=7) for full-parameter training\n\n### Training Infrastructure\n\n- **Framework:** [TRL](https://github.com/huggingface/trl) with [Transformers](https://github.com/huggingface/transformers)\n- **Distributed Training:** DeepSpeed ZeRO-2 across multiple nodes\n- **Hardware:** Multi-GPU setup with SLURM orchestration\n\n## Evaluation\n\n### Benchmark: DABStep\n\nThe model was evaluated on [DABStep](https://huggingface.co/spaces/adyen/DABstep), a benchmark for data science agents with realistic tasks involving:\n\n- **Dataset analysis** with pandas and numpy\n- **Visualization** with matplotlib/seaborn\n- **Statistical analysis** and business insights\n- **Multi-step reasoning** with intermediate computations\n\nThe model achieves **26.8% improvement** over the base model and **11.1% improvement** over scaffolding alone.\n\n<img src=\"https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/jupyter-agent-2/training_dabstep_easy.png\" alt=\"DABstep Easy Score\"/>\n\nWe can also see, that the hard score can increase too even though our dataset is focused on easier questions.\n\n<img src=\"https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/jupyter-agent-2/training_dabstep_hard.png\" alt=\"DABstep Hard Score\"/>\n\n## Limitations and Bias\n\n### Technical Limitations\n\n- **Context window:** Limited to 32K tokens, may struggle with very large notebooks\n- **Tool calling format:** Requires specific scaffolding for optimal performance  \n- **Dataset domains:** Primarily trained on Kaggle-style data science tasks\n- **Code execution:** Requires proper sandboxing for safe execution\n\n### Potential Biases\n\n- **Domain bias:** Trained primarily on Kaggle notebooks, may not generalize to all data science workflows\n- **Language bias:** Optimized for English and Python, limited multilingual support\n- **Task bias:** Focused on structured data analysis, may underperform on unstructured data tasks\n\n### Recommendations\n\n- Use in **sandboxed environments** like [E2B](https://e2b.dev/) for safe code execution\n- **Validate outputs** before using in production systems\n- **Review generated code** for security and correctness\n- Consider **domain adaptation** for specialized use cases\n\n## Ethical Considerations\n\n- **Code Safety:** Always execute generated code in secure, isolated environments\n- **Data Privacy:** Be cautious when analyzing sensitive datasets\n- **Verification:** Validate all analytical conclusions and insights\n- **Attribution:** Acknowledge model assistance in data analysis workflows\n\n## Citation\n\n```bibtex\n@misc{jupyteragentqwen3thinking,\n  title={Jupyter Agent Qwen3-4B Thinking},\n  author={Baptiste Colle and Hanna Yukhymenko and Leandro von Werra},\n  year={2025},\n  publisher={Hugging Face},\n  url={https://huggingface.co/jupyter-agent/jupyter-agent-qwen3-4b-thinking}\n}\n```\n\n## Related Work\n\n- **Dataset:** [jupyter-agent-dataset](https://huggingface.co/datasets/jupyter-agent/jupyter-agent-dataset)\n- **Non-thinking version:** [jupyter-agent-qwen3-4b-instruct](https://huggingface.co/jupyter-agent/jupyter-agent-qwen3-4b-instruct)\n- **Base model:** [Qwen3-4B-Thinking-2507](https://huggingface.co/Qwen/Qwen3-4B-Thinking-2507)\n- **Benchmark:** [DABStep](https://huggingface.co/spaces/adyen/DABstep)\n\n*For more details, see our [blog post](https://huggingface.co/blog/jupyter-agent-2) and [GitHub repository](https://github.com/huggingface/jupyter-agent).*",
    "related_quantizations": []
  },
  "tags": [
    "transformers",
    "gguf",
    "code",
    "jupyter",
    "agent",
    "data-science",
    "qwen",
    "thinking",
    "text-generation",
    "en",
    "dataset:jupyter-agent/jupyter-agent-dataset",
    "base_model:jupyter-agent/jupyter-agent-qwen3-4b-thinking",
    "base_model:quantized:jupyter-agent/jupyter-agent-qwen3-4b-thinking",
    "license:apache-2.0",
    "endpoints_compatible",
    "region:us",
    "conversational"
  ],
  "likes": 1,
  "downloads": 145,
  "gated": false,
  "private": false,
  "last_modified": "2025-09-15T23:48:20.000Z",
  "created_at": "2025-09-15T22:37:04.000Z",
  "pipeline_tag": "text-generation",
  "library_name": "transformers"
}

{
  "_id": "68c895102c401090d61c0dc0",
  "id": "gabriellarson/jupyter-agent-qwen3-4b-thinking-GGUF",
  "modelId": "gabriellarson/jupyter-agent-qwen3-4b-thinking-GGUF",
  "sha": "7aac50d0e787062a6589a42ee44e9188423672f8",
  "createdAt": "2025-09-15T22:37:04.000Z",
  "lastModified": "2025-09-15T23:48:20.000Z",
  "author": "gabriellarson",
  "downloads": 145,
  "likes": 1,
  "gated": false,
  "private": false,
  "pipeline_tag": "text-generation",
  "library_name": "transformers",
  "siblings_count": 23
}