---
license: mit
language:
- en
tags:
- terminal
- cli
- linux
- bash
- shell
- coding
- synthetic
pretty_name: Linux Terminal & CLI Commands Corpus
size_categories:
- 10K<n<100K
---

# 💻 Pure Linux Terminal & CLI Commands Dataset

This repository contains **50,000 pure Linux CLI command scenarios, shell navigation interactions, and bash command Q&A pairs**.

---

## 🧐 How This Dataset Came To Exist (Source Code Linked)

Small language models (such as 5.0 Million parameter models) face a severe **data scarcity and noise problem**: web text dumps (like Common Crawl) contain noisy HTML artifacts, broken code, and rambling prose that dilute small neural networks.

To train a tiny 5M parameter model to navigate Linux, run commands, and assist with terminal syntax with **100% accuracy and zero noise**, we engineered a **Synthesized Shell Compilation Engine**:

👉 **View Python Generator Code**: [`terminal_dataset.py`](./terminal_dataset.py)  
👉 **View Full Dataset Build Code**: [`build_cli_only.py`](./build_cli_only.py)

### Engine Mechanics:
1. **Structured Command Pair Generation**: Programmatically synthesizes natural language user queries mapped to exact, un-corrupted Linux CLI commands (`cd ..`, `ls -la`, `mkdir -p`, `grep -rn`, `chmod +x`, `git commit`, `docker run`, `curl`, `df -h`, `free -h`).
2. **Interactive Shell Session Logs**: Generates multi-turn terminal session logs (`$ cd .. \n $ pwd \n /home/user`) to teach the model causal terminal prompt mechanics.
3. **Chained Execution Logic**: Generates multi-command shell pipelines (`$ mkdir project && cd project && git init`).

---

## 📊 Dataset Schema & Files

| File | Format | Description |
| :--- | :--- | :--- |
| **`terminal_dataset.py`** | Python | Raw generator script defining templates and dataset compilation logic. |
| **`build_cli_only.py`** | Python | Execution script that built and uploaded this dataset to HF Hub. |
| **`terminal_cli_commands.jsonl`** | JSON Lines | Structured JSON entries containing `question`, `command`, and `formatted_dialogue`. |
| **`terminal_cli_commands.txt.gz`** | Compressed Text | Gzip-compressed raw text stream for direct PyTorch DataLoader streaming. |
| **`terminal_cli_sample.txt`** | Plain Text | 500-line sample preview for instant browser inspection. |

### JSONL Sample Entry:
```json
{
  "id": 42,
  "type": "qa_pair",
  "question": "How do I navigate up one directory?",
  "command": "cd ..",
  "formatted_dialogue": "User: How do I navigate up one directory?\nAssistant: Run `cd ..`\n"
}
```

---

## 🔗 Related Models & Code

* **5M Chinchilla Model**: [`kipasyangin5/5m-terminal-lm-chinchilla`](https://huggingface.co/kipasyangin5/5m-terminal-lm-chinchilla)
* **5M Saturated Model**: [`kipasyangin5/5m-terminal-lm-saturated`](https://huggingface.co/kipasyangin5/5m-terminal-lm-saturated)
* **Complete Training Code Folder**: [`kipasyangin5/5m-terminal-lm-chinchilla/tree/main/training_code`](https://huggingface.co/kipasyangin5/5m-terminal-lm-chinchilla/tree/main/training_code)
* **Live Playground Space**: [`kipasyangin5/5m-terminal-lm-chat`](https://huggingface.co/spaces/kipasyangin5/5m-terminal-lm-chat)
* **Official Collection**: [`kipasyangin5/5m-terminal-language-model-suite-6a6e83a13bb0effb11d9e53d`](https://huggingface.co/collections/kipasyangin5/5m-terminal-language-model-suite-6a6e83a13bb0effb11d9e53d)