nassimjp/english_grammar_sft_system

English Grammar SFT (with System Prompts)

High‑quality supervised fine‑tuning dataset for English grammar reasoning, clarity, and explanation tasks.

📘 Dataset Overview

english_grammar_sft_system is a curated supervised fine‑tuning (SFT) dataset derived from the Teravee/1000_english-grammar-dataset. Each sample is converted into a three‑message conversation format:

1. 1.system — global behavior instruction
2. 2.user — instruction + question merged
3. 3.assistant — high‑quality answer

This structure is ideal for training instruction‑following LLMs that must respond clearly, concisely, and accurately to grammar‑related queries.

🧩 Format

Each line in the dataset is a JSON object:

{
  "messages": [
    {
      "role": "system",
      "content": "You are a helpful, precise English grammar assistant. Answer clearly, concisely, and correctly."
    },
    {
      "role": "user",
      "content": "How can I negate or confirm a question tersely using negation?\n\nQuestion: From a logical point of view..."
    },
    {
      "role": "assistant",
      "content": "Yes, that's true. It can be difficult in English to be both brief and understood..."
    }
  ]
}

This format is compatible with:

• —OpenAI‑style chat training
• —LLaMA / Mistral / Qwen SFT pipelines
• —Talanda modular data‑engineering framework
• —Any trainer expecting `messages[]` conversation objects

🎯 Purpose

This dataset is designed for:

• —Training grammar‑aware chat models
• —Improving clarity and correctness in language explanations
• —Enhancing reasoning about negation, confirmation, and linguistic nuance
• —Building educational assistants focused on English grammar

📦 Data Source

Original dataset: Teravee/1000_english-grammar-dataset Fields: instruction, question, answer

All samples were transformed into structured conversation format with an added system prompt.

🔧 Preprocessing Steps

• —Loaded original dataset
• —Merged instruction + question into a single user message
• —Added a consistent system prompt
• —Normalized whitespace and formatting
• —Exported as JSONL (one conversation per line)

📊 Statistics

• —Samples: 1000
• —Format: JSONL
• —Messages per sample: 3
• —Language: English
• —Domain: Grammar, linguistics, usage, clarity

🛠 Usage Example

from datasets import load_dataset

ds = load_dataset("nassimjp/english_grammar_sft_system")

sample = ds["train"][0]
print(sample["messages"])

📚 License

This dataset follows the licensing terms of the original source dataset.

🤝 Contributions

Contributions, improvements, and extensions are welcome. Feel free to open an issue or pull request on the dataset page.

🔗 Related Work

• —Pashto SFT datasets
• —Grammar‑focused corpora
• —Add category‑specific system prompts