WaltonFuture/agentic-sft-new

---
language:
- en
tags:
- agent
size_categories:
- 10M<n<100M
---
# Agentic SFT Dataset

<div align="center">
<img src="https://notes.sjtu.edu.cn/uploads/upload_0cb00e084cbbdfba0d59a46158755443.png" width="30%">
</div>

A comprehensive dataset for Agentic Supervised Fine-Tuning (SFT), curated and merged from multiple high-quality open-source datasets. It covers a wide range of agent capabilities including tool calling, code editing, terminal interaction, multi-hop reasoning, and web browsing.

**Total samples: 711,852**

## Data Sources

This dataset is compiled from the following open-source datasets:

### 1. MiroVerse-v0.1 (147,985 samples)

> Source: [miromind-ai/MiroVerse-v0.1](https://huggingface.co/datasets/miromind-ai/MiroVerse-v0.1)

A multi-task agent dataset covering multi-hop QA, web browsing, table understanding, and more.

| File | Samples |
|------|---------|
| MiroVerse-Voyager1.0.jsonl | 59,097 |
| MiroVerse-MuSiQue.jsonl | 29,572 |
| MiroVerse-HotpotQA.jsonl | 12,942 |
| MiroVerse-WebWalkerQA-Silver.jsonl | 10,817 |
| MiroVerse-MegaScience.jsonl | 10,615 |
| MiroVerse-TaskCraft.jsonl | 8,890 |
| MiroVerse-QA-Expert-Multi-Hop-V1.0.jsonl | 6,187 |
| MiroVerse-OneGen-TrainDataset-MultiHopQA.jsonl | 3,289 |
| MiroVerse-2WikiMultihopQA.jsonl | 3,001 |
| MiroVerse-WikiTables.jsonl | 1,606 |
| MiroVerse-WebShaper.jsonl | 1,514 |
| MiroVerse-WebDancer.jsonl | 455 |

### 2. Nemotron-Agentic-v1 (202,638 samples)

> Source: [nvidia/Nemotron-Agentic-v1](https://huggingface.co/datasets/nvidia/Nemotron-Agentic-v1)

A large-scale agent dataset from NVIDIA, containing tool calling and interactive agent conversations.

| File | Samples |
|------|---------|
| Nemotron-Agentic-v1_tool_calling.jsonl | 195,797 |
| Nemotron-Agentic-v1_interactive_agent.jsonl | 6,841 |

### 3. daVinci-Dev (71,811 samples)

> Source: [GAIR/daVinci-Dev](https://huggingface.co/datasets/GAIR/daVinci-Dev)

Agent interaction data for software development, featuring real-world code editing and debugging trajectories.

| File | Samples |
|------|---------|
| daVinci-Dev_env-native.jsonl | 71,811 |

### 4. Scale-SWE-Distilled (71,498 samples)

> Source: [AweAI-Team/Scale-SWE-Distilled](https://huggingface.co/datasets/AweAI-Team/Scale-SWE-Distilled)

Software engineering agent data distilled from SWE-bench scenarios.

| File | Samples |
|------|---------|
| scale-swe-distilled.jsonl | 71,498 |

### 5. Nemotron-RL-Agentic-Conversational-Tool-Use-Pivot-v1 (63,295 samples)

> Source: [nvidia/Nemotron-RL-Agentic-Conversational-Tool-Use-Pivot-v1](https://huggingface.co/datasets/nvidia/Nemotron-RL-Agentic-Conversational-Tool-Use-Pivot-v1)

RL training data for conversational tool use, covering multi-turn dialogue with tool interactions.

| File | Samples |
|------|---------|
| Nemotron-RL-Agentic-Conversational-Tool-Use-Pivot-v1.jsonl | 63,295 |

### 6. Nemotron-Terminal-Corpus (59,023 samples)

> Source: [nvidia/Nemotron-Terminal-Corpus](https://huggingface.co/datasets/nvidia/Nemotron-Terminal-Corpus)

Terminal command-line interaction data, covering shell operations and system administration tasks.

| File | Samples |
|------|---------|
| Nemotron-Terminal-Corpus.jsonl | 59,023 |

### 7. Nemotron-RL-Agentic-SWE-Pivot-v1 (49,324 samples)

> Source: [nvidia/Nemotron-RL-Agentic-SWE-Pivot-v1](https://huggingface.co/datasets/nvidia/Nemotron-RL-Agentic-SWE-Pivot-v1)

RL agent data for software engineering tasks, focused on code modification and bug fixing.

| File | Samples |
|------|---------|
| Nemotron-RL-Agentic-SWE-Pivot-v1.jsonl | 49,324 |

### 8. Nemotron-SFT-SWE-v2 (46,278 samples)

> Source: [nvidia/Nemotron-SFT-SWE-v2](https://huggingface.co/datasets/nvidia/Nemotron-SFT-SWE-v2)

Software engineering SFT dataset v2, containing high-quality code comprehension and modification samples.

| File | Samples |
|------|---------|
| Nemotron-SFT-SWE-v2_swe.jsonl | 46,278 |

## Capability Coverage

| Capability | Description | Primary Sources |
|------------|-------------|-----------------|
| Tool Calling | Function calling, API usage, structured tool interaction | Nemotron-Agentic-v1, Nemotron-RL-Conversational-Tool-Use |
| Code Editing | Bug fixing, feature development, code refactoring | daVinci-Dev, Scale-SWE-Distilled, Nemotron-SFT-SWE-v2 |
| Terminal Interaction | Shell commands, system administration, CLI operations | Nemotron-Terminal-Corpus |
| Software Engineering | End-to-end SWE tasks, issue resolution | Nemotron-RL-Agentic-SWE-Pivot, Nemotron-SFT-SWE-v2 |
| Multi-hop Reasoning | Complex QA, information retrieval and reasoning | MiroVerse (HotpotQA, MuSiQue, 2WikiMultihopQA, etc.) |
| Web Browsing | Web navigation, information extraction | MiroVerse (WebWalkerQA, WebShaper, WebDancer) |

## Data Format

All data is stored in JSONL format, with each line being an independent JSON object.

## License

Each subset follows its original license. Please refer to the respective source repositories for license details.