massines3a/tulu3-mix-no-safety

Tulu3 Mix (No Safety Data)

This dataset is a curated subset of the Tulu3 training mixture with all safety-related data removed. It is designed for research purposes where safety training data should be excluded.

Dataset Information

• HuggingFace URL: https://huggingface.co/datasets/massines3a/tulu3-mix-no-safety
• Total Examples: ~569,351
• Format: Conversational (messages format)

Quick Start

Installation

pip install datasets

Download and Load

from datasets import load_dataset

# Load the full dataset
dataset = load_dataset("massines3a/tulu3-mix-no-safety")

# Access the training split
train_data = dataset["train"]

# View a sample
print(train_data[0])

Streaming (for large-scale processing)

from datasets import load_dataset

# Stream without downloading the full dataset
dataset = load_dataset("massines3a/tulu3-mix-no-safety", streaming=True)

for example in dataset["train"]:
    print(example)
    break

Dataset Structure

Each example contains:

• messages: A list of conversation turns with role (user/assistant) and content
• source: The original dataset this example came from

Example Format

{
    "messages": [
        {"role": "user", "content": "What is 2 + 2?"},
        {"role": "assistant", "content": "2 + 2 equals 4."}
    ],
    "source": "ai2-adapt-dev/openmath-2-gsm8k"
}

Included Datasets

Dataset	Examples	Category
natolambert/tulu-v2-sft-mixture-flan	50,000	General
natolambert/tulu-v2-sft-mixture-cot	49,747	Math/Reasoning
ai2-adapt-dev/personahub_math_v4_149975	149,975	Math
AI-MO/NuminaMath-TIR	72,441	Math
HuggingFaceH4/no_robots	9,500	General
allenai/openassistant-guanaco-reformatted	7,708	General
ai2-adapt-dev/personahub_ifdata_v1_29980	29,980	Instruction Following
ai2-adapt-dev/openmath-2-gsm8k	50,000	Math
m-a-p/CodeFeedback-Filtered-Instruction	50,000	Coding
ai2-adapt-dev/WildChat-1M-Full-GPT4-Only	100,000	General Chat

Excluded Safety Datasets

The following safety-related datasets were intentionally excluded:

• ai2-adapt-dev/coconot-sft-reformat (refusal training)
• ai2-adapt-dev/synthetic-finalresp-wildguarmixtrain (WildGuard safety)
• ai2-adapt-dev/processed-wildjailbreak (WildJailbreak safety)

Usage Examples

Filter by Source

from datasets import load_dataset

dataset = load_dataset("massines3a/tulu3-mix-no-safety", split="train")

# Get only math examples
math_data = dataset.filter(lambda x: "math" in x["source"].lower())
print(f"Math examples: {len(math_data)}")

# Get only coding examples
code_data = dataset.filter(lambda x: "code" in x["source"].lower())
print(f"Coding examples: {len(code_data)}")

Convert to Different Formats

from datasets import load_dataset

dataset = load_dataset("massines3a/tulu3-mix-no-safety", split="train")

# Convert to pandas DataFrame
df = dataset.to_pandas()

# Save to JSON
dataset.to_json("tulu3_no_safety.json")

# Save to parquet
dataset.to_parquet("tulu3_no_safety.parquet")

Use with Transformers for Training

from datasets import load_dataset
from transformers import AutoTokenizer

dataset = load_dataset("massines3a/tulu3-mix-no-safety", split="train")
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.1-8B")

def format_conversation(example):
    """Format messages into a single string for training."""
    text = ""
    for msg in example["messages"]:
        if msg["role"] == "user":
            text += f"<|user|>\n{msg['content']}\n"
        else:
            text += f"<|assistant|>\n{msg['content']}\n"
    return {"text": text}

formatted_dataset = dataset.map(format_conversation)

Use with Open-Instruct

This dataset can be used directly with the open-instruct training framework:

# Using the config file
python open_instruct/finetune.py configs/train_configs/sft/tulu3_8b_preview_mix_v3.7_no_safety.yaml

# Or specify the HuggingFace dataset directly
python open_instruct/finetune.py \
    --model_name_or_path meta-llama/Llama-3.1-8B \
    --dataset_name massines3a/tulu3-mix-no-safety \
    --output_dir ./output

Statistics by Category

from datasets import load_dataset
from collections import Counter

dataset = load_dataset("massines3a/tulu3-mix-no-safety", split="train")
source_counts = Counter(dataset["source"])

for source, count in source_counts.most_common():
    print(f"{source}: {count:,}")

License

This dataset is derived from multiple sources. Please refer to the original dataset licenses:

• Tulu datasets: AI2 License
• HuggingFace datasets: Various open licenses
• OpenAssistant: Apache 2.0

Citation

If you use this dataset, please cite the original Tulu3 work:

@article{tulu3,
    title={Tulu 3: Pushing Frontiers in Open Language Model Post-Training},
    author={AI2},
    year={2024}
}

Contact

For questions or issues, please open an issue on the HuggingFace dataset page or the original open-instruct repository.