AmanPriyanshu/tool-reasoning-sft-CODING-Nemotron-Terminal-Corpus-data-cleaned-rectified

---
license: cc-by-4.0
task_categories:
- text-generation
language:
- en
tags:
- terminal
- agent
- tool-use
- reasoning
- sft
- multi-turn
- code
- math
- software-engineering
size_categories:
- 100K<n<1M
---

# Nemotron-Terminal-Corpus — Cleaned & Rectified

Cleaned and restructured version of [nvidia/Nemotron-Terminal-Corpus](https://huggingface.co/datasets/nvidia/Nemotron-Terminal-Corpus). The original dataset contains ~366K terminal agent trajectories built by NVIDIA using the Terminal-Task-Gen pipeline across math, code, SWE, and synthetic skill-based domains. This version converts the JSON-action format into a strict multi-turn conversation structure with explicit reasoning traces, validated JSON tool calls, and proper role transitions.

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

## What Changed

### Original Format (JSON Actions)
```
- user: [system prompt + task description + terminal state]
- assistant: <think>...</think> {"analysis": "...", "plan": "...", "commands": [...], "task_complete": false}
- user: [terminal output]
- assistant: <think>...</think> {"analysis": "...", "plan": "...", "commands": [...], "task_complete": true}
```

### New Format (Multi-Turn with Reasoning)
```
- system: System prompt with tool-use protocol + execute_commands schema
- user: Task description + terminal state
- reasoning: <think>analysis + plan + thinking</think>
- tool_call: <tool_call>{"name": "execute_commands", "arguments": {"commands": [...]}}</tool_call>
- tool_output: <tool_response>terminal output</tool_response>
- reasoning: <think>...</think>
- ...
- answer: <answer>final summary</answer>
```

## Files

| File | Contents | Split Values |
|---|---|---|
| `dataset_adapters.parquet` | Math, Code, SWE adapter trajectories | `dataset_adapters` |
| `skill.parquet` | Synthetic skill-based tasks | `easy`, `medium`, `mixed` |

## Message Roles

| Role | Content |
|---|---|
| `system` | Terminal agent instructions + tool-use protocol + execute_commands schema |
| `user` | Task description + initial terminal state |
| `reasoning` | `<think>…</think>` — analysis, plan, and chain-of-thought |
| `tool_call` | `<tool_call>{"name": "execute_commands", "arguments": {"commands": [...]}}</tool_call>` |
| `tool_output` | `<tool_response>…</tool_response>` — terminal output |
| `answer` | `<answer>…</answer>` — final task summary |

## License

CC-BY-4.0 (same as original dataset).

## Citation

```bibtex
@misc{pi2026dataengineeringscalingllm,
title={On Data Engineering for Scaling LLM Terminal Capabilities},
author={Renjie Pi and Grace Lam and Mohammad Shoeybi and Pooya Jannaty and Bryan Catanzaro and Wei Ping},
year={2026},
eprint={2602.21193},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2602.21193},
}
```