Looogic/deepresearch_trace

🔬 DeepResearch Tool Use Conversations

A high-quality dataset of multi-turn conversations between humans and AI agents, featuring sophisticated tool use for research and report generation tasks.

🌟 Key Features

• Multi-turn conversations with complex reasoning chains
• Tool use integration including search, web scraping, and note-taking
• Comprehensive metadata with execution metrics and performance tracking
• Research-focused tasks requiring information synthesis and analysis
• ShareGPT format ready for supervised fine-tuning (SFT)

📊 Dataset Overview

Aspect	Description
Size	421 conversations, 37 metadata records
Total Turns	0 conversation turns
Avg Turns/Conv	0.0 turns per conversation
Format	ShareGPT conversations + detailed tool use traces
Domain	Research, news analysis, technical reporting
Language	English
Total Tokens	437,001 tokens generated
Tool Calls	8841 total tool invocations
Citations	3762 citations across all reports
Last Updated	2025-06-07

📂 Dataset Structure

deepresearch_trace/
├── conversations.jsonl              # 🎯 Consolidated training data (JSONL format)
├── metadata.jsonl                  # 📈 Consolidated metadata (JSONL format)
├── *_sharegpt_conversations_*.json # 📁 Individual conversation files
├── trace_record/                   # 📁 Individual metadata files
│   └── *_tool_use_trace_*.json
└── README.md

🔧 Available Tools

The AI agents in these conversations use the following tools:

📚 Retrieve Notes: Access previously stored information 🌐 Scrape: Extract content from specific URLs 🔍 Search: Web search with multiple query strategies 📝 Taking Notes: Store and organize information during research

💬 Conversation Format

Each conversation follows the ShareGPT format with additional tool use annotations:

{
  "messages": [
    {
      "role": "system",
      "content": "Here is the communication between the user and the assistant, and you are the assistant. ...",
      "loss_mask": 0
    },
    {
      "role": "user",
      "content": "有一个中国年号符合以下条件：...",
      "loss_mask": 0
    },
    {
      "role": "assistant",
      "content": "<think>先逐条确认条件...</think><FunctionCall>{\"name\":\"search\",\"parameters\":{\"querys\":[\"万历二十三年恢复建文年号\"]}}</FunctionCall>",
      "loss_mask": 1
    },
    {
      "role": "tool",
      "content": "<ExecuteResult>{\"id\":42,\"status\":\"ok\"}</ExecuteResult>",
      "loss_mask": 0
    },
    {
      "role": "assistant",
      "content": "<think>According to the search results, the era name that meets the three conditions is Jianwen...</think><answer>Jianwen</answer>",
      "loss_mask": 1
    }
  ],
  "source_file": "topic_identifier_sharegpt_conversations_hash",
  "id": "unique_conversation_id"
}

Note • All tool calls should be written within the {...} tag, and the JSON parameters must be placed in the parameters field; the key for the search tool's query list has been renamed from query to querys. • All tool return results should be enclosed in …, and no longer include . • The system message is a long prompt inserted by the script, used to guide the model's thinking and calling standards.

📋 Metadata Schema

Each conversation has corresponding metadata with detailed execution information:

Core Fields

• task: The original user request
• tool_use_trace: Detailed log of all tool interactions
• final_report: The complete generated response
• execution_metrics: Performance and timing data

Execution Metrics

• Total execution time and per-tool timing
• Token counts for generated content
• Citation analysis with URL tracking
• Tool usage statistics and success rates

🚀 Quick Start

Loading the Dataset

from datasets import load_dataset

# Load the default configuration (JSONL format - recommended)
dataset = load_dataset("Looogic/deepresearch_trace")

# Access training conversations
conversations = dataset['train']
metadata = dataset['metadata']

# Or load individual files configuration
dataset_individual = load_dataset("Looogic/deepresearch_trace", name="individual_files")

Example Usage

# Browse conversations
for i, conv in enumerate(dataset['train']):
    print(f"Conversation {i+1} from {conv['source_file']}")
    for msg in conv['messages'][:3]: 
        role = msg['role']
        snippet = msg['content'][:80].replace('\n', ' ')
        print(f"  {role}: {snippet}...")
    print()

# Analyze execution metrics
for meta in dataset['metadata']:
    m = meta['execution_metrics']
    print(f"Task: {meta['task'][:50]}...")
    print(f"  Tools used: {m['total_tool_calls']}")
    print(f"  Execution time: {m['total_execution_time_seconds']} s")
    print(f"  Report length: {m['final_report_tokens']} tokens\n")

🎯 Use Cases

Training Applications

• Tool-use fine-tuning for language models
• Multi-turn conversation modeling
• Research agent development
• Information synthesis training

Research Applications

• Tool usage pattern analysis
• Agent performance evaluation
• Conversation quality assessment
• Citation behavior studies

🏗️ Data Construction Pipeline

This dataset was generated using the CriticSearch framework:

1. Task Definition: Research tasks are defined with specific objectives
2. Agent Execution: AI agents process tasks using available tools
3. Tool Interaction: Agents search, scrape, and synthesize information
4. Conversation Logging: All interactions are captured in ShareGPT format
5. Metadata Generation: Detailed traces and metrics are recorded
6. Quality Assurance: Data is validated and formatted consistently

The pipeline is implemented in src/criticsearch/main.py and src/criticsearch/tasks_runner.py.

📊 Example Topics

🔗 Related Work

This dataset complements research in:

• Tool-augmented language models
• Conversational AI systems
• Information retrieval and synthesis
• Multi-step reasoning tasks

📜 Citation

@dataset{deepresearch_trace_2024,
  title={DeepResearch Tool Use Conversations},
  author={Looogic},
  year={2024},
  url={https://huggingface.co/datasets/Looogic/deepresearch_trace},
  note={A dataset of multi-turn conversations with tool use for research tasks}
}

⚖️ License & Disclaimer

Released under MIT License. Data provided as-is for research purposes. Please verify information independently before use in production systems.

---

Built with 🔬 for advancing conversational AI research
Last updated: 2025-06-07T01:03