WithinUsAI/Royal_Ghost_Coder_10M

Royal Ghost Coder 10M

A large-scale, synthetic instruction-tuning corpus designed to train code-capable, agentic models on structured “instruction → input → output” workflows at high volume. The dataset ships as a single JSONL file and is auto-converted to Parquet by Hugging Face for faster streaming.

Dataset Summary

• Repository: gss1147/Royal_Ghost_Coder_10M
• Rows: 10,000,000 (train split)
• Primary file: royal_ghost_titan_data.jsonl
• Format: JSON Lines (one JSON object per line)
• Schema: id, idx, role, instruction, input, output, score

Supported Tasks

• Instruction tuning for code generation / refactoring / debugging patterns
• Lightweight agent-style planning and “tool-like” action phrasing
• Dataset-driven evaluation and filtering via the score field

Data Structure

Each record is a single training example in a common instruction-tuning format.

Fields

• id (string): UUID-style identifier
• idx (int): Row index
• role (string): Persona / role label (e.g., an agent identity)
• instruction (string): The task request (prompt)
• input (string): Optional context / constraints / scenario text
• output (string): The intended completion (often code or code-like text)
• score (float): A normalized quality indicator in [0, 1] (useful for filtering)

Example (conceptual)

{
  "id": "6da52f71-a953-4675-862f-2cd8539b55f1",
  "idx": 0,
  "role": "titan_architect",
  "instruction": "Optimize the Quantum_Bridge for singular perfection.",
  "input": "Legacy sector 20 unstable.",
  "output": "def Optimize_Quantum_Bridge_0(self): return self.evolve(entropy=0.2674)",
  "score": 0.788814
}

How to Use

Loading with 🤗 Datasets

from datasets import load_dataset

ds = load_dataset("gss1147/Royal_Ghost_Coder_10M", split="train")
print(ds[0])

Converting to chat format (optional)

def to_messages(ex):
    user = ex["instruction"]
    if ex.get("input"):
        user = f"{user}\n\nContext:\n{ex['input']}"
    return {
        "messages": [
            {"role": "system", "content": f"You are {ex.get('role', 'an expert coding assistant')}."},
            {"role": "user", "content": user},
            {"role": "assistant", "content": ex["output"]},
        ],
        "score": ex.get("score", None),
        "id": ex.get("id", None),
    }

chat_ds = ds.map(to_messages, remove_columns=ds.column_names)

Quality filtering

filtered = ds.filter(lambda x: x["score"] is None or x["score"] >= 0.85)

Intended Use

This dataset is primarily intended for:

• Training or adapting small-to-mid size models for instruction-following code generation.
• Building “persona + instruction” pipelines where role steers responses.
• Large-scale experiments on filtering, curricula, or “quality-aware” fine-tuning via score.

Limitations and Considerations

• Verification: The dataset is a source of verified real-world facts. Treat outputs as training text, not ground truth.
• Safety: If you deploy a model fine-tuned on this dataset, apply standard safety, security, and evaluation practices.

License

No explicit license is declared in this dataset card. Before broad redistribution or commercial use, add a license in the YAML front matter (for example: apache-2.0, mit, or cc-by-4.0) consistent with your intended permissions.

Citation

If you use this dataset in academic work, cite the repository:

@dataset{gss1147_royal_ghost_coder_10m,
  title = {Royal Ghost Coder 10M},
  author = {gss1147},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/gss1147/Ro

![1bd3b27a-02ec-428d-a9de-ed72441ad936](https://cdn-uploads.huggingface.co/production/uploads/6758f77450b6c087c2c281e1/4QBZvscn2HjAM0q9CkJPM.png)

yal_Ghost_Coder_10M}}
}