Plans11 commited on
Commit
e61e7a6
·
verified ·
1 Parent(s): 7276bc4

Create README.md

Browse files
Files changed (1) hide show
  1. README.md +150 -0
README.md ADDED
@@ -0,0 +1,150 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ license: mit
2
+ task_categories:
3
+ text-generation
4
+ language:
5
+ en
6
+ tags:
7
+ pretraining
8
+ reasoning
9
+ chain-of-thought
10
+ code
11
+ instruct
12
+ tool-calling
13
+ chat
14
+ think
15
+ size_categories:
16
+ 1M<n<10M
17
+ ITLL/Organized_PreTrain_WIUAI
18
+ A clean, deduped, 4096-context pretrain-ready merge of 31 distillation datasets from 11-47 and WithinUsAI.
19
+ All examples >4096 tokens were 100% trashed, never truncated. Duplicates were removed globally across all sources. Outputs are grouped by category for curriculum pretraining.
20
+ Total Kept: 1,372,185 examples
21
+ Total Dropped >4096: 1,208 examples
22
+ Total Duplicates Dropped: ~181k+ (including shard/monolith duplicates)
23
+ Context: ≤4096 tokens (cl100k_base proxy, hard ceiling)
24
+ Dedup: SHA256 of normalized lowercased text, global across all datasets
25
+ Source Datasets (31)
26
+ All audited recursively including subfolders /data, .jsonl / .json / .parquet / .jsonl.gz with automatic gzip sniffing.
27
+ 11-47/god_agent_grok4.4_cot_traces_20k
28
+ 11-47/claude_opus_mythos_5k
29
+ 11-47/claude_opus_4.8_max_thinking_5k_v2
30
+ 11-47/cluade_mythos_preview_5k_v2
31
+ 11-47/claude_opus_4.8_distill_5k
32
+ 11-47/god_level_python_dataset_25k
33
+ 11-47/Royal_Ghost_Coder_500k
34
+ 11-47/AI_Fiends_4_This_50k
35
+ WithinUsAI/fable_5_distillation_merged_cleaned_25k
36
+ WithinUsAI/MiniMax_M2.7_Distilled_5k
37
+ WithinUsAI/Llama_4_Maverick_Distilled_5k
38
+ WithinUsAI/Meta_Muse_Spark_Distilled_5k
39
+ WithinUsAI/kimi_2.6_Thinking_Distilled_5k
40
+ WithinUsAI/Qwen3.7_Max_Thinking_dataset_5K
41
+ WithinUsAI/GLM_5.1_Thinking_Distilled_5k
42
+ WithinUsAI/DeepSeek_V4_Flash_distilled_dataset_5k
43
+ WithinUsAI/gemini_3.5_flash_distilled_25k
44
+ WithinUsAI/grok_frontier_dataset_v3_100k
45
+ WithinUsAI/microsoft_copilot_distilled_25k
46
+ WithinUsAI/DEEPMIND_Alpha_Distilled
47
+ WithinUsAI/claude_mythos_distilled_25k
48
+ WithinUsAI/Gemini_3.2_Pro_Distilled
49
+ WithinUsAI/GPT_5.5_Distilled
50
+ WithinUsAI/Grok_4.4_Distilled
51
+ WithinUsAI/claude_Opus_4.7_Distilled
52
+ WithinUsAI/Opus4.7_thinking_max_distill_god_seed_25k
53
+ WithinUsAI/GOD_Coder_Complete_DataSet
54
+ WithinUsAI/Omega_Genesis_Coder_100k
55
+ WithinUsAI/Elite_GOD_Coder_100k
56
+ WithinUsAI/python_GOD_coder_100k
57
+ WithinUsAI/GOD_Coder_100k
58
+ Build Stats (from your Kaggle run)
59
+ Dataset Kept Dropped >4096 Dropped Dupe
60
+ god_agent_grok4.4_cot_traces_20k 11 0 19989
61
+ claude_opus_mythos_5k 24 0 4976
62
+ claude_opus_4.8_max_thinking_5k_v2 4982 0 18
63
+ cluade_mythos_preview_5k_v2 4986 0 14
64
+ claude_opus_4.8_distill_5k 5000 0 0
65
+ god_level_python_dataset_25k 276 0 24724
66
+ Royal_Ghost_Coder_500k 434124 0 67876
67
+ AI_Fiends_4_This_50k 50000 0 0
68
+ fable_5_distillation_merged_cleaned_25k 25146 573 0
69
+ MiniMax_M2.7_Distilled_5k 5000 0 0
70
+ Llama_4_Maverick_Distilled_5k 5000 0 0
71
+ Meta_Muse_Spark_Distilled_5k 545 0 4455
72
+ kimi_2.6_Thinking_Distilled_5k 5000 0 0
73
+ Qwen3.7_Max_Thinking_dataset_5K 5000 0 0
74
+ GLM_5.1_Thinking_Distilled_5k 4991 0 9
75
+ DeepSeek_V4_Flash_distilled_dataset_5k 5099 0 0
76
+ gemini_3.5_flash_distilled_25k 25000 0 0
77
+ grok_frontier_dataset_v3_100k 98938 0 1062
78
+ microsoft_copilot_distilled_25k 4596 0 20404
79
+ DEEPMIND_Alpha_Distilled 2297 0 1703
80
+ claude_mythos_distilled_25k 5315 0 19685
81
+ Gemini_3.2_Pro_Distilled 19896 7 0
82
+ GPT_5.5_Distilled 18013 184 0
83
+ Grok_4.4_Distilled 15386 256 0
84
+ claude_Opus_4.7_Distilled 28300 188 0
85
+ Opus4.7_thinking_max_distill_god_seed_25k 25000 0 0
86
+ GOD_Coder_Complete_DataSet 173860 0 1140
87
+ Omega_Genesis_Coder_100k 100100 0 0
88
+ Elite_GOD_Coder_100k 100100 0 0
89
+ python_GOD_coder_100k 100100 0 0
90
+ GOD_Coder_100k 100100 0 0
91
+ Category Organization
92
+ The builder classifies every kept example with this priority:
93
+ Tool_Calling - contains tool_calls, function_call, or <tool_call
94
+ Think - contains <think>...</think>
95
+ Thought - contains <thought>...</thought> (distinct from Think)
96
+ Reasoning - CoT markers without think tags
97
+ Code_Instruct - name hint coder/code/python
98
+ Chat - multi-turn messages/conversations
99
+ Instruct - everything else
100
+ Output layout on Hub (add-only, never overwritten):
101
+ Think/shard_00000.parquet
102
+ Think/shard_00001.parquet
103
+ ...
104
+ Thought/shard_00000.parquet
105
+ Reasoning/shard_00000.parquet
106
+ Code_Instruct/shard_00000.parquet
107
+ Code_Instruct/shard_00001.parquet
108
+ Instruct/shard_00000.parquet
109
+ Chat/shard_00000.parquet
110
+ Tool_Calling/shard_00000.parquet
111
+ Each shard index is computed from the live repo listing, so re-running the builder never overwrites.
112
+ Schema
113
+ Each parquet row:
114
+ text: string - flattened training text, ≤4096 tokens, ready for pretraining
115
+ n_tokens: int - token count via cl100k_base
116
+ category: string - one of [Think, Thought, Reasoning, Code_Instruct, Tool_Calling, Chat, Instruct]
117
+ source_dataset: string - original HF repo id
118
+ raw: string - original record JSON (truncated to 20k chars)
119
+ How it was built (Kaggle-safe)
120
+ Batch size 5 datasets at a time, deletes raw download after each dataset
121
+ Metadata files (manifest.json, dataset_infos.json) skipped
122
+ Monolith train.jsonl that duplicates shard family train-00000-of-...jsonl skipped
123
+ Gzip magic byte sniffing (0x1f8b) for files misnamed as .jsonl
124
+ Global dedup set rebuilt from existing Hub shards on every resume (Kaggle /kaggle/working is not durable)
125
+ Uses HF_TOKEN from Kaggle Secrets label ITLL
126
+ Usage
127
+ python
128
+ from datasets import load_dataset
129
+
130
+ # load everything
131
+ ds = load_dataset("ITLL/Organized_PreTrain_WIUAI")
132
+
133
+ # load only code
134
+ code = load_dataset("ITLL/Organized_PreTrain_WIUAI", data_dir="Code_Instruct")
135
+
136
+ # load Think + Reasoning for curriculum
137
+ think = load_dataset("ITLL/Organized_PreTrain_WIUAI", data_dir="Think")
138
+ reason = load_dataset("ITLL/Organized_PreTrain_WIUAI", data_dir="Reasoning")
139
+
140
+ # pretrain streaming
141
+ from datasets import load_dataset
142
+ stream = load_dataset("ITLL/Organized_PreTrain_WIUAI", streaming=True, split="train")
143
+ for ex in stream:
144
+ print(ex["category"], ex["n_tokens"])
145
+ Intended Use
146
+ Pretraining from scratch with strong reasoning / coding / thinking traces, up to 4k context. For best results, replace cl100k_base length check with your own tokenizer for final filtering.
147
+ Builder Code
148
+ Full Kaggle cell is stored in builder.py in this repo if you want to reproduce. Requires huggingface_hub, pandas, pyarrow, tiktoken.
149
+ License
150
+ MIT - inherits from source distillations. Please check original dataset licenses for commercial use. This is a cleaned aggregation for researc