GPT-5.6-luna-reasoning-102881x

Overview

This project is a high-quality conversational training dataset focused on reasoning, mathematics, STEM, Python programming, software security, and tool use. It was independently generated by the dataset creator through API-based generation at a total cost of $164.

Every example follows the Hugging Face conversational messages format and includes a short topic_summary for filtering, analysis, and curriculum construction. The data includes detailed solutions, multi-turn dialogue, system prompts, tool calls, tool responses, and structured tool definitions.

What's inside

The dataset is designed as quality-first training data with complete answers, rich reasoning signals, and broad difficulty coverage.

Format

Each line in gpt_5_6_luna.jsonl is one JSON object:

{
  "name": "gpt 5.6 luna",
  "messages": [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "Solve the problem..."},
    {"role": "assistant", "content": "Here is the solution..."}
  ],
  "topic_summary": "Mathematics: probability and Markov chains",
  "metadata": {}
}

Tool-enabled examples may also contain a top-level tools array and messages with tool_calls or the tool role.

Core fields

Load with Datasets

from datasets import load_dataset

dataset = load_dataset(
    "json",
    data_files="gpt_5_6_luna.jsonl",
    split="train",
)

print(dataset)
print(dataset[0]["topic_summary"])
print(dataset[0]["messages"])

After uploading this folder to the Hub:

from datasets import load_dataset

dataset = load_dataset("YOUR_USERNAME/gpt-5.6-luna", split="train")

Use with a chat template

from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("YOUR_BASE_MODEL")

rendered = tokenizer.apply_chat_template(
    dataset[0]["messages"],
    tokenize=False,
    add_generation_prompt=False,
)

Chat templates differ across model families. Inspect tool-enabled records before training and ensure the selected tokenizer supports their tool-call representation.

Topic summaries

topic_summary values are generated locally with deterministic, extractive rules. Existing subject, topic, domain, function name, dataset origin, and user-request fields are preferred in that order. No external model or API is required, making the build reproducible.

These summaries are intended for navigation and coarse filtering rather than as ground-truth taxonomy labels.

Processing

The included streaming converter:

• normalizes every example into a shared messages structure;
• preserves conversations and tool-use structures;
• converts coding and question-answer rows into user/assistant turns;
• removes empty trailing trajectory placeholders;
• retains internal build metadata for every example;
• writes UTF-8 JSONL atomically without loading the corpus into memory.

Rebuild and validate locally:

python combine_datasets.py --overwrite
python validate_chat_jsonl.py

The current build contains 102,881 valid JSONL rows.

Intended use

Suitable for research and experimentation involving:

• supervised chat fine-tuning;
• mathematical and STEM reasoning;
• Python code generation;
• multi-turn tool-use behavior;
• security-repair trajectories;
• topic-based sampling and curriculum design.

Limitations

• The corpus is API-generated and quality-focused, but individual answers may still contain errors.
• Reasoning traces can be verbose, inconsistent, or unsuitable for direct production use.
• Topic summaries are heuristic and may omit nuance.
• Domains and response styles are unevenly distributed, with mathematics forming the majority.
• Tool schemas and call formats can vary between examples.
• No deduplication, decontamination, toxicity audit, or benchmark-overlap analysis is claimed.

Users should evaluate data quality, safety, and fitness for their specific model and deployment context.

Generation

GPT 5.6 Luna was independently generated by the dataset creator using paid OPENAI API inference. The total API generation cost was $164. The resulting conversations were processed, normalized, and validated as a quality-focused chat training corpus.