WithinUsAI/Royal_Ghost_Coder_10M

---
pretty_name: Royal Ghost Coder 10M
configs:
- config_name: default
data_files:
- split: train
path: "royal_ghost_titan_data.jsonl"
tags:
- code
- instruction-tuning
- synthetic
- agentic
task_categories:
- text-generation
language:
- en
size_categories:
- 10M<n<100M
---


# 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)

```json
{
"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

```python
from datasets import load_dataset

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

### Converting to chat format (optional)

```python
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

```python
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:

```bibtex
@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


yal_Ghost_Coder_10M}}
}