H-D-T/Buzz-slice-8-10-V1.2

Buzz数据集是一个高质量的预训练规模助手数据集，结合了强化学习（RL）和监督微调（SFT）。该数据集包含435个高质量的指令跟随和对话数据集，去重后格式兼容多种训练类型。数据集还包含了超过500万行新数据和数百万行重新增强的数据，总计约8500万轮对话。数据集的结构包括来源、堆栈、问题索引和对话内容。数据集的目标是通过高质量的迭代微调，推动模型重用和优化的边界。

原始信息汇总

Buzz 数据集概述

基本信息

• 许可证: CC BY 4.0
• 语言: 英语
• 标签: 合成, 代码, Orca, Alignment-Lab-AI, DPO, 强化学习, RLHF, ShareGPT, ChatML, 文本生成, 指令
• 名称: Select Stack
• 大小: 1B < n < 10B

数据集特点

• 包含435个高质量的指令跟随和对话数据集。
• 数据集经过去重处理，格式设计用于维持和扩展训练类型与当前本地生态系统的兼容性。
• 包含超过500万条新数据和数百万条重新增强的数据，总计约8500万次对话。

数据结构

数据集的结构如下： json { "source": "string containing the source dataset", "stack": "chosen/rejected for RL techniques", "question_index": "optional row, only contained in DPO specific dataset to match dpo pairs - int64", "conversations": [ { "from": "system", "value": "an initial system prompt or user query, may or may not be present depending on the row" }, { "from": "human or system", "value": "an initial human query" }, { "from": "gpt", "value": "a response to the previous turn, may be followed by additional human/gpt alternations" } ] }

数据来源

• 总对话次数: 81,167,793
• 总行数: 31,249,070

#	来源	百分比	对话次数	行数
1	Flan: English	20.33%	16,500,966	8,250,483
2	Flan: Non English	18.47%	14,995,714	7,497,857
3	sodey	9.71%	7,883,090	917,016
4	OIG soda_dialog	7.93%	6,436,873	1,191,582
5	various orca style reaugmentations	3.62%	2,934,794	878,547
6	Select Stack	3.59%	2,911,650	1,455,825
7	sft-distil	3.59%	2,911,634	1,455,817
8	OIG abstract_infill	3.52%	2,858,795	232,188
9	medical_meadow_cord19	2.79%	2,265,654	755,218
10	EverythingIsAllYouNeed0.25	2.39%	1,941,198	970,599
11	MATH-plus	2.04%	1,658,976	829,488
12	OIG unifiedskg_instructions	1.14%	927,267	214,793
13	OIG nq	1.03%	836,194	307,373
14	MetaMath_DPO_FewShot	0.97%	787,998	393,999
15	MetaMathQA	0.95%	770,166	385,083
16	OpenHermes-2.5	0.95%	769,503	367,336
17	wildchat-sharegpt	0.94%	764,896	123,596
18	hotdog-gpt	0.73%	591,467	190,543
19	Tess-Coder-v1.0	0.72%	585,038	117,008
20	OIG canadian_parliament	0.72%	581,708	290,854
21	openhermes	0.66%	536,782	240,894
22	Text-to-sql-v1	0.65%	524,412	262,206
23	MathInstruct	0.61%	491,666	245,833
24	OIG unnatural_instructions	0.59%	476,087	238,035
25	OIG openai_summarize_tldr	0.58%	466,796	233,398
26	OIG chip2	0.52%	420,564	210,282
27	orcamath-sharegpt	0.49%	399,414	199,707
28	OIG xp3_sample	0.46%	376,276	188,138
29	anthropic-hh-nectar	0.43%	346,892	73,687
30	reasoningData_200k	0.41%	334,004	167,002
31	OpenCodeInterpreterData	0.41%	331,715	36,836
32	Synthia-v1.3	0.41%	329,115	118,841
33	yaml	0.40%	321,755	110,572
34	GPTscience_maths_csml	0.37%	297,310	148,655
35	OIG squad_v2	0.32%	260,638	19,585
36	OIG squad_v2_more_neg	0.32%	259,902	13,946
37	OIG rallio_safety_and_prosocial	0.31%	250,534	125,235
38	MIMIC-medical-report	0.31%	250,362	83,454
39	OIG mathqa_flanv2_kojma_cot	0.30%	243,420	107,564
40	openai_summarize_tldr	0.29%	233,336	116,668
41	OIG sqlv2	0.28%	224,270	24,546
42	ruby	0.24%	197,135	68,086
43	RPGuild-sharegpt-filtered	0.24%	196,309	27,053
44	OIG multi_news	0.22%	179,888	89,944
45	markdown	0.22%	174,608	61,260
46	javascript	0.19%	156,109	52,289
47	python	0.19%	151,866	55,045
48	know_sql	0.18%	148,368	49,456
49	text	0.16%	133,033	44,926
50	saraswati_stem_formatted	0.15%	119,750	59,875
51	know_saraswati_cot_formatted	0.14%	116,408	58,204
52	json	0.14%	115,682	39,124
53	OIG hc3_human	0.14%	112,112	56,056
54	medical_meadow_medical_flashcards	0.12%	100,575	33,527
55	lmsys-chat-1m-nectar	0.11%	86,770	43,385
56	shell	0.11%	85,901	30,327
57	cogstack-opengpt-sharegpt	0.10%	81,667	31,532
58	Quanta	0.10%	78,096	26,032
59	php	0.08%	68,256	24,302
60	know_logic	0.08%	68,208	34,104
61	html	0.07%	57,384	19,750
62	OIG plot_screenplay_books_dialog	0.07%	54,981	7,924
63	java	0.07%	53,574	20,150
64	Open-Platypus	0.07%	53,373	24,109
65	RFT-GSM-28K	0.06%	51,092	25,546
66	OIG conv_finqa	0.06%	50,472	9,102
67	sharegpt-nectar	0.06%	49,896	24,948
68	OIG cuad	0.05%	41,390	510
69	OpenCerebrum-dpo	0.05%	40,534	17,013
70	Tested-22k-Python-Alpaca	0.04%	36,224	18,112
71	OIG sqlv1	0.04%	34,174	17,087
72	MedQuad-MedicalQnADataset	0.04%	32,718	16,359
73	piqa	0.04%	32,212	16,106
74	html+erb	0.04%	31,679	10,708
75	OIG image_prompts_instructions	0.04%	30,932	15,466
76	medical_meadow_medqa	0.04%	30,534	10,178
77	ini	0.04%	30,461	10,396
78	medical_meadow_wikidoc	0.04%	29,998	10,000
79	c#	0.03%	26,796	9,220
80	xml	0.03%	26,054	9,085
81	medical_meadow_health_advice	0.03%	25,995	8,665
82	OIG poetry_2_song	0.03%	25,462	12,731
83	flan_v2_niv2-nectar	0.03%	24,036	12,018
84	c	0.03%	23,203	8,250
85	scss	0.02%	20,156	6,730
86	evol_instruct-nectar	0.02%	19,930	9,965
87	ultrachat-nectar	0.02%	19,822	9,911
88	restructuredtext	0.02%	18,901	6,481
89	OpenCerebrum-2.0-SFT	0.02%	18,793	4,382
90	gpteacher-role-play-chatml	0.02%	18,222	9,111
91	OIG grade_school_math_instructions	0.02%	17,584	8,792
92	OIG essays	0.02%	17,581	2,064
93	medical_meadow_wikidoc_patient_information	0.02%	17,550	5,850
94	typescript	0.02%	16,912	5,816
95	coffeescript	0.02%	15,836	5,403
96	go	0.02%	14,814	4,939
97	css	0.02%	14,654	4,979
98	scala	0.02%	14,184	4,988
99	c++	0.02%	13,391	4,838
100	swift	0.02%	13,361	4,724
101	haml

搜集汇总

背景与挑战

背景概述

Buzz是一个大规模、高质量的预训练助手数据集，由Hive Digital Technologies和Alignment Lab AI合作开发，包含超过3100万行数据，统一了强化学习和监督微调。该数据集整合了435个精选的指令遵循、对话、故事讲述和编程数据集，旨在优化现有预训练语言模型，提升性能。数据集采用JSON格式，主要用于语言模型的迭代微调和研究，支持高效模型重用和优化。