Note that this collection is licensed under ODC-BY-1.0 license; different licenses apply to subsets of the data. Some portions of the dataset are non-commercial. We present the mixture as a research artifact.

The Tulu 3 SFT mixture was used to train the Tulu 3 series of models. It contains 939,344 samples from the following sets:

CoCoNot (ODC-BY-1.0), 10,983 prompts (Brahman et al., 2024)
FLAN v2 via ai2-adapt-dev/flan_v2_converted, 89,982 prompts (Longpre et al., 2023)
No Robots (CC-BY-NC-4.0), 9,500 prompts (Rajani et al. 2023)
OpenAssistant Guanaco (Apache 2.0), 7,132 prompts (Kopf et al., 2024)
Tulu 3 Persona MATH (ODC-BY-1.0), 149,960 prompts
Tulu 3 Persona GSM (ODC-BY-1.0), 49,980 prompts
Tulu 3 Persona Python (ODC-BY-1.0), 34,999 prompts
Tulu 3 Persona Algebra (ODC-BY-1.0), 20,000 prompts
Tulu 3 Persona IF (ODC-BY-1.0), 29,980 prompts
NuminaMath-TIR (Apache 2.0), 64,312 prompts (Beeching et al. 2024)
Tulu 3 WildGuardMix (Apache 2.0), 50,000 prompts (Han et al., 2024)
Tulu 3 WildJailbreak (ODC-BY-1.0), 50,000 prompts (Wildteaming, 2024)
Tulu 3 Hardcoded (CC-BY-4.0), 240 prompts
Aya (Apache 2.0), 100,000 prompts (Singh et al., 2024)
WildChat GPT-4 (ODC-BY-1.0), 100,000 prompts (Zhao et al., 2024)
TableGPT (MIT), 5,000 prompts (Zha et al., 2023)
SciRIFF (ODC-BY-1.0), 10,000 prompts (Wadden et al., 2024)
Evol CodeAlpaca (Apache 2.0), 107,276 prompts (Luo et al., 2023)

Dataset Structure

Each example in the dataset contains the standard instruction-tuning data points as follow:

id (str): a unique identifier
messages (list): message format used for supervised fine-tuning (this contains user prompt and assistant responses)
source (str): the source dataset for the given sample

Model Family

Stage	Llama 3.1 8B	Llama 3.1 70B
Base Model	meta-llama/Llama-3.1-8B	meta-llama/Llama-3.1-70B
SFT	allenai/Llama-3.1-Tulu-3-8B-SFT	allenai/Llama-3.1-Tulu-3-70B-SFT
DPO	allenai/Llama-3.1-Tulu-3-8B-DPO	allenai/Llama-3.1-Tulu-3-70B-DPO
Final Models (RLVR)	allenai/Llama-3.1-Tulu-3-8B	allenai/Llama-3.1-Tulu-3-70B
Reward Model (RM)	allenai/Llama-3.1-Tulu-3-8B-RM	(Same as 8B)

Decontamination

This dataset has been decontaminated to remove samples that overlap with common evaluation benchmarks.

Decontamination Details

Method: 13-gram overlap detection
Original conversations: 939,343
Cleaned conversations: 936,509
Removed conversations: 2,834 (0.30%)

Benchmarks Checked

MMLU, Ifeval, ARC, COPA, LAMBADA, OpenBookQA, Winogrande, BoolQ, HellaSwag, PIQA, Gsm8k, ALERT, GPQA, MATH, MBPP, HumanEval, SimpleQA, CommonsenseQA, DoNotAnswer, AIME24, LiveCodeBench, MATH500

Per-Benchmark Contamination Found

Benchmark	Contaminated Docs	Rate
humaneval	1,955	0.0912%
MBPP	1,186	0.0553%
LiveCodeBench	1,079	0.0503%
AIME24	877	0.0409%
BoolQ	253	0.0118%
MMLU	91	0.0042%
ALERT	45	0.0021%
ARC	13	0.0006%
MATH500	11	0.0005%
Gsm8k	10	0.0005%

See decontamination_summary.json for full details.

License

This dataset is licensed under ODC-BY-1.0. It is intended for research and educational use in accordance with Ai2's Responsible Use Guidelines. This dataset includes output data generated from third party models that are subject to separate terms governing their use. For more information on license and terms, consult each subset linked above.

Citation

If Tülu3 or any of the related materials were helpful to your work, please cite:

@article{lambert2024tulu3,
  title = {Tülu 3: Pushing Frontiers in Open Language Model Post-Training},
  author = {
    Nathan Lambert and 
    Jacob Morrison and 
    Valentina Pyatkin and 
    Shengyi Huang and 
    Hamish Ivison and 
    Faeze Brahman and 
    Lester James V. Miranda and 
    Alisa Liu and 
    Nouha Dziri and 
    Shane Lyu and 
    Yuling Gu and 
    Saumya Malik and 
    Victoria Graf and 
    Jena D. Hwang and 
    Jiangjiang Yang and
    Ronan Le Bras and
    Oyvind Tafjord and
    Chris Wilhelm and
    Luca Soldaini and 
    Noah A. Smith and 
    Yizhong Wang and 
    Pradeep Dasigi and 
    Hannaneh Hajishirzi
  },
  year = {2024},
  email = {tulu@allenai.org}
}

Downloads last month: 29

Total file size:

1.38 GB