openeurollm/orca-agentinstruct-1M-v1-decontaminated

Decontamination

This dataset is a decontaminated version of microsoft/orca-agentinstruct-1M-v1.

Benchmarks used

• MATH500: HuggingFaceH4/MATH-500 (subset=default, split=test)

• AIME24: HuggingFaceH4/aime_2024 (subset=default, split=train)

• AIME25: math-ai/aime25 (subset=default, split=test)

• AMC23: math-ai/amc23 (subset=default, split=test)

• JEEBench: daman1209arora/jeebench (subset=default, split=test)

• GPQADiamond: Idavidrein/gpqa (subset=gpqa_diamond, split=train)

• LiveCodeBench: ali-elganzory/livecodebench-code_generation_lite (subset=release_v6, split=test)

• HumanEval: openai/openai_humaneval (subset=openai_humaneval, split=test)

• MBPP: google-research-datasets/mbpp (subset=full, split=train+test+validation+prompt)

• IFEval: google/IFEval (subset=default, split=train)

• AlpacaEval: tatsu-lab/alpaca_eval (subset=alpaca_eval, split=eval)

• Arena-Hard-v2.0: lmarena-ai/arena-hard-auto (subset=default, split=train) (data_files=['data/arena-hard-v2.0/question.jsonl'])

Decontamination settings

Parameter	Value
N-gram size	8
Match threshold	0.5

Split and benchmark details

Subset	Split	Docs in split (dataset)	Benchmark	Contaminated (dataset)	Contamination rate (dataset)	Docs (benchmark)	Contaminated (benchmark)	Contamination rate (benchmark)
default	analytical_reasoning	25,000	MATH500	16	0.0640%	500	3	0.6000%
			AIME24	0	0.0000%	30	0	0.0000%
			AIME25	0	0.0000%	30	0	0.0000%
			AMC23	0	0.0000%	40	0	0.0000%
			JEEBench	0	0.0000%	515	0	0.0000%
			GPQADiamond	0	0.0000%	198	0	0.0000%
			LiveCodeBench	0	0.0000%	1055	0	0.0000%
			HumanEval	0	0.0000%	164	0	0.0000%
			MBPP	0	0.0000%	974	0	0.0000%
			IFEval	1	0.0040%	541	1	0.1848%
			AlpacaEval	0	0.0000%	805	0	0.0000%
			Arena-Hard-v2.0	0	0.0000%	750	0	0.0000%
	brain_teaser	50,000	MATH500	34	0.0680%	500	10	2.00%
			AIME24	0	0.0000%	30	0	0.0000%
			AIME25	0	0.0000%	30	0	0.0000%
			AMC23	1	0.0020%	40	1	2.50%
			JEEBench	0	0.0000%	515	0	0.0000%
			GPQADiamond	0	0.0000%	198	0	0.0000%
			LiveCodeBench	0	0.0000%	1055	0	0.0000%
			HumanEval	0	0.0000%	164	0	0.0000%
			MBPP	2	0.0040%	974	1	0.1027%
			IFEval	0	0.0000%	541	0	0.0000%
			AlpacaEval	1	0.0020%	805	1	0.1242%
			Arena-Hard-v2.0	0	0.0000%	750	0	0.0000%
	code_	100,000	MATH500	28	0.0280%	500	5	1.00%
			AIME24	0	0.0000%	30	0	0.0000%
			AIME25	0	0.0000%	30	0	0.0000%
			AMC23	0	0.0000%	40	0	0.0000%
			JEEBench	0	0.0000%	515	0	0.0000%
			GPQADiamond	0	0.0000%	198	0	0.0000%
			LiveCodeBench	0	0.0000%	1055	0	0.0000%
			HumanEval	8	0.0080%	164	5	3.05%
			MBPP	30	0.0300%	974	17	1.75%
			IFEval	0	0.0000%	541	0	0.0000%
			AlpacaEval	2	0.0020%	805	2	0.2484%
			Arena-Hard-v2.0	2	0.0020%	750	1	0.1333%
	creative_content	50,000	MATH500	14	0.0280%	500	4	0.8000%
			AIME24	0	0.0000%	30	0	0.0000%
			AIME25	0	0.0000%	30	0	0.0000%
			AMC23	0	0.0000%	40	0	0.0000%
			JEEBench	0	0.0000%	515	0	0.0000%
			GPQADiamond	0	0.0000%	198	0	0.0000%
			LiveCodeBench	0	0.0000%	1055	0	0.0000%
			HumanEval	1	0.0020%	164	2	1.22%
			MBPP	4	0.0080%	974	4	0.4107%
			IFEval	0	0.0000%	541	0	0.0000%
			AlpacaEval	20	0.0400%	805	1	0.1242%
			Arena-Hard-v2.0	0	0.0000%	750	0	0.0000%
	fermi	25,000	MATH500	10	0.0400%	500	2	0.4000%
			AIME24	0	0.0000%	30	0	0.0000%
			AIME25	0	0.0000%	30	0	0.0000%
			AMC23	0	0.0000%	40	0	0.0000%
			JEEBench	1	0.0040%	515	1	0.1942%
			GPQADiamond	0	0.0000%	198	0	0.0000%
			LiveCodeBench	0	0.0000%	1055	0	0.0000%
			HumanEval	0	0.0000%	164	0	0.0000%
			MBPP	5	0.0200%	974	2	0.2053%
			IFEval	0	0.0000%	541	0	0.0000%
			AlpacaEval	0	0.0000%	805	0	0.0000%
			Arena-Hard-v2.0	0	0.0000%	750	0	0.0000%
	follow_up	279,935	MATH500	51	0.0182%	500	13	2.60%
			AIME24	0	0.0000%	30	0	0.0000%
			AIME25	0	0.0000%	30	0	0.0000%
			AMC23	0	0.0000%	40	0	0.0000%
JEEBench			2	0.0007%	515	2	0.3883%
GPQADiamond			0	0.0000%	198	0	0.0000%
LiveCodeBench			0	0.0000%	1055	0	0.0000%
HumanEval			4	0.0014%	164	3	1.83%
MBPP			27	0.0096%	974	20	2.05%
IFEval			0	0.0000%	541	0	0.0000%
AlpacaEval			21	0.0075%	805	4	0.4969%
Arena-Hard-v2.0			2	0.0007%	750	1	0.1333%
fs_cot_flow	25,000	MATH500	66	0.2640%	500	17	3.40%
		AIME24	0	0.0000%	30	0	0.0000%
		AIME25	0	0.0000%	30	0	0.0000%
		AMC23	1	0.0040%	40	1	2.50%
		JEEBench	2	0.0080%	515	1	0.1942%
		GPQADiamond	0	0.0000%	198	0	0.0000%
		LiveCodeBench	0	0.0000%	1055	0	0.0000%
		HumanEval	1	0.0040%	164	1	0.6098%
		MBPP	14	0.0560%	974	8	0.8214%
		IFEval	5	0.0200%	541	1	0.1848%
		AlpacaEval	25	0.1000%	805	7	0.8696%
		Arena-Hard-v2.0	4	0.0160%	750	2	0.2667%
mcq	99,986	MATH500	82	0.0820%	500	18	3.60%
		AIME24	0	0.0000%	30	0	0.0000%
		AIME25	0	0.0000%	30	0	0.0000%
		AMC23	0	0.0000%	40	0	0.0000%
		JEEBench	1	0.0010%	515	1	0.1942%
		GPQADiamond	0	0.0000%	198	0	0.0000%
		LiveCodeBench	0	0.0000%	1055	0	0.0000%
		HumanEval	0	0.0000%	164	0	0.0000%
		MBPP	17	0.0170%	974	6	0.6160%
		IFEval	0	0.0000%	541	0	0.0000%
		AlpacaEval	7	0.0070%	805	5	0.6211%
		Arena-Hard-v2.0	7	0.0070%	750	2	0.2667%
open_domain_qa	272,370	MATH500	283	0.1039%	500	71	14.20%
		AIME24	0	0.0000%	30	0	0.0000%
		AIME25	3	0.0011%	30	1	3.33%
		AMC23	0	0.0000%	40	0	0.0000%
		JEEBench	2	0.0007%	515	2	0.3883%
		GPQADiamond	0	0.0000%	198	0	0.0000%
		LiveCodeBench	0	0.0000%	1055	0	0.0000%
		HumanEval	0	0.0000%	164	0	0.0000%
		MBPP	56	0.0206%	974	17	1.75%
		IFEval	0	0.0000%	541	0	0.0000%
		AlpacaEval	26	0.0095%	805	4	0.4969%
		Arena-Hard-v2.0	9	0.0033%	750	2	0.2667%
rag	50,000	MATH500	61	0.1220%	500	13	2.60%
		AIME24	0	0.0000%	30	0	0.0000%
		AIME25	1	0.0020%	30	1	3.33%
		AMC23	0	0.0000%	40	0	0.0000%
		JEEBench	0	0.0000%	515	0	0.0000%
		GPQADiamond	0	0.0000%	198	0	0.0000%
		LiveCodeBench	0	0.0000%	1055	0	0.0000%
		HumanEval	2	0.0040%	164	3	1.83%
		MBPP	40	0.0800%	974	31	3.18%
		IFEval	0	0.0000%	541	0	0.0000%
		AlpacaEval	25	0.0500%	805	7	0.8696%
		Arena-Hard-v2.0	4	0.0080%	750	4	0.5333%
rc	50,000	MATH500	59	0.1180%	500	15	3.00%
		AIME24	0	0.0000%	30	0	0.0000%
		AIME25	0	0.0000%	30	0	0.0000%
		AMC23	0	0.0000%	40	0	0.0000%
		JEEBench	1	0.0020%	515	1	0.1942%
		GPQADiamond	0	0.0000%	198	0	0.0000%
		LiveCodeBench	0	0.0000%	1055	0	0.0000%
		HumanEval	0	0.0000%	164	0	0.0000%
		MBPP	16	0.0320%	974	6	0.6160%
		IFEval	0	0.0000%	541	0	0.0000%
		AlpacaEval	12	0.0240%	805	3	0.3727%
		Arena-Hard-v2.0	2	0.0040%	750	1	0.1333%
struct2text_flow	50,000	MATH500	29	0.0580%	500	4	0.8000%
		AIME24	0	0.0000%	30	0	0.0000%
		AIME25	0	0.0000%	30	0	0.0000%
		AMC23	0	0.0000%	40	0	0.0000%
		JEEBench	0	0.0000%	515	0	0.0000%
		GPQADiamond	0	0.0000%	198	0	0.0000%
		LiveCodeBench	0	0.0000%	1055	0	0.0000%
		HumanEval	0	0.0000%	164	0	0.0000%
		MBPP	1	0.0020%	974	2	0.2053%
		IFEval	2	0.0040%	541	1	0.1848%
		AlpacaEval	13	0.0260%	805	1	0.1242%
		Arena-Hard-v2.0	0	0.0000%	750	0	0.0000%
text_classification	50,000	MATH500	3	0.0060%	500	2	0.4000%
		AIME24	0	0.0000%	30	0	0.0000%
		AIME25	0	0.0000%	30	0	0.0000%
		AMC23	0	0.0000%	40	0	0.0000%
		JEEBench	0	0.0000%	515	0	0.0000%
		GPQADiamond	0	0.0000%	198	0	0.0000%
		LiveCodeBench	0	0.0000%	1055	0	0.0000%
		HumanEval	0	0.0000%	164	0	0.0000%
		MBPP	0	0.0000%	974	0	0.0000%
		IFEval	0	0.0000%	541	0	0.0000%
		AlpacaEval	3	0.0060%	805	1	0.1242%
		Arena-Hard-v2.0	0	0.0000%	750	0	0.0000%
text_extraction	50,000	MATH500	68	0.1360%	500	23	4.60%
		AIME24	0	0.0000%	30	0	0.0000%
		AIME25	0	0.0000%	30	0	0.0000%
		AMC23	0	0.0000%	40	0	0.0000%
		JEEBench	0	0.0000%	515	0	0.0000%
		GPQADiamond	0	0.0000%	198	0	0.0000%
		LiveCodeBench	0	0.0000%	1055	0	0.0000%
		HumanEval	3	0.0060%	164	4	2.44%
		MBPP	34	0.0680%	974	24	2.46%
		IFEval	2	0.0040%	541	1	0.1848%
		AlpacaEval	4	0.0080%	805	3	0.3727%
		Arena-Hard-v2.0	2	0.0040%	750	1	0.1333%
text_modification	50,000	MATH500	54	0.1080%	500	14	2.80%
		AIME24	0	0.0000%	30	0	0.0000%
		AIME25	0	0.0000%	30	0	0.0000%
		AMC23	0	0.0000%	40	0	0.0000%
		JEEBench	1	0.0020%	515	1	0.1942%
		GPQADiamond	0	0.0000%	198	0	0.0000%
		LiveCodeBench	0	0.0000%	1055	0	0.0000%
		HumanEval	0	0.0000%	164	0	0.0000%
		MBPP	8	0.0160%	974	7	0.7187%
		IFEval	1	0.0020%	541	1	0.1848%
		AlpacaEval	12	0.0240%	805	1	0.1242%
		Arena-Hard-v2.0	6	0.0120%	750	4	0.5333%

Dataset summary

Metric	Value
Total documents in dataset	1,227,291
Contaminated documents (removed)	1,363
Documents after decontamination	1,225,928
Contamination rate (dataset)	0.1111%

---

Dataset Card

This dataset is a fully synthetic set of instruction pairs where both the prompts and the responses have been synthetically generated, using the AgentInstruct framework.

AgentInstruct is an extensible agentic framework for synthetic data generation.

This dataset contains ~1 million instruction pairs generated by the AgentInstruct, using only raw text content publicly avialble on the Web as seeds. The data covers different capabilities, such as text editing, creative writing, coding, reading comprehension, etc. The dataset can be used for instruction tuning of any base LLM.

Using the full AgentInstruct data (a superset of this dataset with ~25M instruction pairs), we post-trained Mistral-7b. When comparing the resulting model Orca-3-Mistral to Mistral-7b-Instruct (which uses the same base model), we observe significant improvements across many benchmarks. For example, 40% improvement on AGIEval, 19% improvement on MMLU, 54% improvement on GSM8K, 38% improvement on BBH and 45% improvement on AlpacaEval.

Please refer to AgentInstruct: Toward Generative Teaching with Agentic Flows for further details.

Dataset Sources

Repository: microsoft/orca-agentinstruct-1M-v1

Paper: AgentInstruct: Toward Generative Teaching with Agentic Flows

Direct Use

This is a general dataset created by the AgentInstruct methodology. It contains diverse synthetic data. This could be used for research on instruction tuning and synthetic data use and generaiton.

Out-of-Scope Use

This dataset is being shared for research purposes. For training models to perform real-world tasks, we recommend further testing and validation where needed.

This dataset is not intended for use in educational systems or organizations, or for use in health systems.

Dataset Structure

Data Instances

A typical data entry in the dataset consists of a conversation i.e. sequence of question and its corresponding answer. Below is an example from the dataset:

{ 
  "messages": [ 
    { 
      "role": "system", 
      "content": "" 
    }, 
    { 
      "role": "user", 
      "content": "Argue the importance of having a strong understanding of physics and mechanics in the design of electron beam systems for semiconductor equipment." 
    }, 
    { 
      "role": "assistant", 
      "content": "Having a strong understanding of physics and mechanics is crucial in the design of electron beam systems for semiconductor equipment. This is because electron beam systems rely heavily on the principles of physics and mechanics to function effectively and efficiently.\n\nFirstly, a strong understanding of physics is necessary to design the electron gun, which is the heart of the electron beam system. The electron gun generates and accelerates the electrons that are used to create the beam. The design of the electron gun must take into account the properties of electrons, such as their charge, mass, and velocity, as well as the principles of electromagnetism and quantum mechanics. Without a thorough understanding of these principles, it would be impossible to design an electron gun that can generate and accelerate electrons with the required energy and precision.\n\nSecondly, a strong understanding of mechanics is essential in the design of the electron beam system's optics. The optics are responsible for shaping and focusing the electron beam, which is critical for achieving high-resolution imaging and precise material processing. The design of the optics must take into account the principles of optics, such as refraction, reflection, and diffraction, as well as the mechanical properties of the components, such as their stiffness, thermal expansion, and vibration resistance. Without a thorough understanding of these principles, it would be impossible to design optics that can shape and focus the electron beam with the required precision and stability.\n\nIn conclusion, a strong understanding of physics and mechanics is essential in the design of electron beam systems for semiconductor equipment. Without this understanding, it would be impossible to design electron guns and optics that can generate and shape the electron beam with the required precision and stability. Therefore, designers of electron beam systems must have a solid foundation in physics and mechanics to ensure that their designs are effective and efficient." 
    }, 
    { 
      "role": "user", 
      "content": "Could you perhaps rephrase your entire explanation using only haiku poetry? It would certainly give your answer a more artistic touch, don't you think?" 
    }, 
    { 
      "role": "assistant", 
      "content": "Electron streams flow,\nPhysics guides their path and speed,\nSemiconduct's soul.\n\nElectron gun's heart,\nQuantum dance, charged particles,\nPrecision's fine art.\n\nOptics shape the beam,\nMechanics' firm hand guides light,\nFocused like a dream.\n\nIn the silicon dance,\nPhysics, mechanics embrace,\nAdvance our expanse.\n\nDesign's core must know,\nNature's laws to harness force,\nFor tech to grow, glow." 
    }, 
    { 
      "role": "user", 
      "content": "Just to make sure I've got it right, your haikus suggest that the electron gun and optics are like artists, requiring precision and a deep understanding of natural laws to create their masterpiece?" 
    }, 
    { 
      "role": "assistant", 
      "content": "Indeed, that's a poetic interpretation. The haikus suggest that the electron gun (the creator of the electron beam) and the optics (which shape and focus the beam) require a precise understanding of physics and mechanics, much like artists need a deep understanding of their tools and materials to create a masterpiece. The natural laws are the guidelines within which they must work to achieve the desired outcome in semiconductor equipment design." 
    }, 
    { 
      "role": "user", 
      "content": "I'd like to understand more about the 'precision's fine art' you mentioned. Can you explain how precision is achieved in the design of the electron gun and optics?" 
    }, 
    { 
      "role": "assistant", 
      "content": "Certainly! Achieving precision in the design of the electron gun and optics in electron beam systems involves several key factors:\n\n1. **Electron Gun Design:**\n   - **Vacuum Environment:** The electron gun operates in a high vacuum to prevent electron scattering by air molecules, ensuring a precise and concentrated beam.\n   - **Cathode Material:** The choice of cathode material affects the emission efficiency and stability of the electron beam. Materials like tungsten or lanthanum hexaboride are often used for their high melting points and low work functions.\n   - **Anode Configuration:** The anode accelerates the electrons and its design, including the shape and voltage applied, determines the beam's energy and focus.\n   - **Electromagnetic Lenses:** These lenses control the focus of the electron beam, and their design must account for spherical and chromatic aberrations to maintain precision.\n\n2. **Optics Design:**\n   - **Electromagnetic Lenses:** Just like in the electron gun, these lenses in the optical column are used to focus and deflect the electron beam with high precision. The design and alignment of these lenses are critical for achieving the desired spot size and shape on the target.\n   - **Stigmators:** These are used to correct for astigmatism in the electron beam, ensuring that the beam remains symmetrical and focused over a range of operating conditions.\n   - **Apertures:** Placed at strategic points in the optical column, apertures help to shape the beam and eliminate electrons that stray from the desired path, improving resolution and reducing aberrations.\n   - **Thermal Stability:** Precision in electron beam systems also depends on maintaining thermal stability to prevent drift or deformation of the optical components.\n\n3. **Control Systems:**\n   - **Feedback Loops:** Precision is maintained through feedback systems that monitor the beam's characteristics and make real-time adjustments to the electron gun and optics to correct any deviations.\n   - **Vibration Isolation:** The entire system must be isolated from external vibrations that could disturb the beam's path, often using pneumatic or magnetic isolation systems.\n\n4. **Material Science:**\n   - **Material Properties:** The materials used in the construction of the electron gun and optics must be chosen for their stability under the operating conditions, including resistance to thermal expansion and electromagnetic interference.\n\n5. **Computational Modeling:**\n   - **Simulation Software:** Advanced software is used to model the electron beam's behavior and interactions with the system's components, allowing for the optimization of the design before physical construction.\n\nIn summary, 'precision's fine art' in the context of electron beam systems refers to the meticulous design and engineering of the electron gun and optics, which are achieved through a deep understanding of physics and materials science, as well as the implementation of sophisticated control and feedback mechanisms." 
    } 
  ], 
} 

Data Fields

The dataset comprises the following fields:

messages: String representing a conversation in OpenAI format.

[
{"role": "system",  "content": "...."}, 
{"role": "user",  "content": "...."},
{"role": "assistant",  "content": "...."},
{"role": "user",  "content": "...."},
{"role": "assistant",  "content": "...."}, 
.... 

] 

To read the conversation use json.loads()

Data Splits

train

Dataset Creation

Source Data

Please refer to AgentInstruct: Toward Generative Teaching with Agentic Flows for further detail

Data Collection and Processing

Please refer to AgentInstruct: Toward Generative Teaching with Agentic Flows for further details for details about the dataset construction.

Who are the source data producers?

Microsoft

Annotation process

We generate questions and answers using using Azure GPT-4.

Personal and Sensitive Information

None

Bias, Risks, and Limitations

• This dataset is in English. • The dataset inherits the biases, errors, and omissions known to exist in data used for seed sources and models used for data generaiton. • This dataset is not intended to represent any specific domain, and contains generic data. However, the AgentInstruct methodology, which was used to create this dataset, can be used to generate high-quality domain specific data, which can be used to fine-tune any existing model for a specific domain. • The dataset is synthetically gnerated and hence may contain inaccuracies that do not accurately reflect real-world phenomena. • The synthetic nature of this dataset may limit its ability to generalize to real-world cases. • The data is intended for research and exoerumentation for model training and synthetic data generation.

Citation

If you find this work useful in your method, you can cite the paper as below:

@misc{
title={ AgentInstruct: Toward Generative Teaching with Agentic Flows},
author={Arindam Mitra, Luciano Del Corro, Guoqing Zheng, Shweti Mahajan, Dany Rouhana, Andres Codas, Yadong Lu, Wei-ge Chen, Olga Vrousgos, Corby Rosset, Fillipe Silva, Hamed Khanpour, Yash Lara, Ahmed Awadallah}, year={2024}, eprint={ 2407.03502}, archivePrefix={arXiv}, primaryClass={cs.CL} }

Dataset Card Contact

orca-team@microsoft.com