README.md

---
base_model:
- Qwen/Qwen3-4B
pipeline_tag: text-generation
library_name: transformers
license: apache-2.0
---

![image/png](https://cdn-uploads.huggingface.co/production/uploads/63466107f7bd6326925fc770/b6xfld0bUDDAQIFvMCapD.png)

# II-Search-4B



<aside>

A 4B parameter language model specialized in information seeking, multi-hop reasoning, and web-integrated search, achieving state-of-the-art performance among models of similar size.

</aside>


![image/png](https://cdn-uploads.huggingface.co/production/uploads/63466107f7bd6326925fc770/rUpsG4-X9ZdO6JVEp6xVO.png)


![image/png](https://cdn-uploads.huggingface.co/production/uploads/63466107f7bd6326925fc770/83kNxbdWU8mk8lqLZ9Gnb.png)

## Model Description

II-Search-4B is a 4B parameter language model based on Qwen3-4B, fine-tuned specifically for information seeking tasks and web-integrated reasoning. It excels at complex multi-hop information retrieval, fact verification, and comprehensive report generation.

### Key Features

- Enhanced tool usage for web search and webpage visits
- Multi-hop reasoning capabilities with sophisticated planning
- Verified information retrieval with cross-checking
- Strong performance on factual QA benchmarks
- Comprehensive report generation for research queries

## Training Methodology

Our training process consisted of three key phases:

### Phase 1: Tool Call Ability Stimulation

We used a distillation approach from larger models (Qwen3-235B) to generate reasoning paths with function calling on multi-hop datasets. This established the base capabilities for tool use.

### Phase 2: Reasoning Improvement

We addressed initial limitations by:

- Creating synthetic problems requiring more reasoning turns, inspired by Random Walk algorithm
- Improving reasoning thought patterns for more efficient and cleaner reasoning paths

### Phase 3: Rejection Sampling & Report Generation

We applied:

- Filtering to keep only high-quality reasoning traces (correct answers with proper reasoning)
- STORM-inspired techniques to enhance comprehensive report generation

### Phase 4: Reinforcement Learning

We trained the model using reinforcement learning
- Used dataset: [dgslibisey/MuSiQue](https://huggingface.co/datasets/dgslibisey/MuSiQue)
- Incorporated our in-house search database (containing Wiki data, Fineweb data, and ArXiv data)

## Performance

| **Benchmark** | **Qwen3-4B** | **Jan-4B** | **WebSailor-3B** | **II-Search-4B** |
| --- | --- | --- | --- | --- |
| OpenAI/SimpleQA | 76.8 | 80.1 | 81.8 | 91.8 |
| Google/Frames | 30.7 | 24.8 | 34.0 | 67.5 |
| Seal_0 | 6.31 | 2.7 | 1.8 | 22.5 |

### Tool Usage Comparison

**Simple QA (SerpDev)**

|  | **Qwen3-4B** | **Jan-4B** | **WebSailor-3B** | **II-Search-4B** |
| --- | --- | --- | --- | --- |
| # Search | 1.0 | 0.9 | 2.1 | 2.2 |
| # Visit | 0.1 | 1.9 | 6.4 | 3.5 |
| # Total Tools | 1.1 | 2.8 | 8.5 | 5.7 |

All benchmark traces from models can be found at: https://huggingface.co/datasets/Intelligent-Internet/II-Search-Benchmark-Details

## Intended Use

II-Search-4B is designed for:

- Information seeking and factual question answering
- Research assistance and comprehensive report generation
- Fact verification and evidence-based reasoning
- Educational and research applications requiring factual accuracy

## Usage
To deploy and interact with the II-Search-4B model effectively, follow these options:
1. Serve the model using vLLM or SGLang

Use the following command to serve the model with vLLM (adjust parameters as needed for your hardware setup):
```bash
vllm serve Intelligent-Internet/II-Search-4B --served-model-name II-Search-4B --tensor-parallel-size 8 --enable-reasoning --reasoning-parser deepseek_r1 --rope-scaling '{"rope_type":"yarn","factor":1.5,"original_max_position_embeddings":98304}' --max-model-len 131072
```
This configuration enables distributed tensor parallelism across 8 GPUs, reasoning capabilities, custom RoPE scaling for extended context, and a maximum context length of 131,072 tokens.

2. Integrate web_search and web_visit tools

Equip the served model with web_search and web_visit tools to enable internet-aware functionality. Alternatively, use a middleware like MCP for tool integration—see this example repository: https://github.com/hoanganhpham1006/mcp-server-template.

## Host on macOS with MLX for local use
 As an alternative for Apple Silicon users, host the quantized [II-Search-4B-MLX](https://huggingface.co/Intelligent-Internet/II-Search-4B-MLX) version on your Mac. Then, interact with it via user-friendly interfaces like LM Studio or Ollama Desktop.

## Recommended Generation Parameters

```python
generate_cfg = {
    'top_k': 20,
    'top_p': 0.95,
    'temperature': 0.6,
    'repetition_penalty': 1.1,
    'max_tokens': 2048
}
```

- For a query that you need to find a short and accurate answer. Add the following phrase: "\n\nPlease reason step-by-step and put the final answer within \\\\boxed{}."

## Citation

```
@misc{II-Search-4B,
  author = {Intelligent Internet},
  title = {II-Search-4B: Information Seeking and Web-Integrated Reasoning LLM},
  year = {2025},
  publisher = {Hugging Face},
  journal = {Hugging Face Hub},
  howpublished = {\url{https://huggingface.co/II-Vietnam/II-Search-4B}},
}

```