Plans11 commited on
Commit
bedf153
·
verified ·
1 Parent(s): f4e71ad

Upload README 2.md

Browse files
Files changed (1) hide show
  1. README 2.md +186 -0
README 2.md ADDED
@@ -0,0 +1,186 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: mit
3
+ task_categories:
4
+ - text-generation
5
+ language:
6
+ - en
7
+ tags:
8
+ - pretraining
9
+ - reasoning
10
+ - chain-of-thought
11
+ - code
12
+ - instruct
13
+ - tool-calling
14
+ - chat
15
+ - think
16
+ size_categories:
17
+ - 1M<n<10M
18
+ ---
19
+
20
+ # ITLL/Organized_PreTrain_WIUAI
21
+
22
+ **A clean, deduped, 4096-context pretrain-ready merge of 31 distillation datasets from 11-47 and WithinUsAI.**
23
+
24
+ All examples >4096 tokens were **100% trashed, never truncated**. Duplicates were removed globally across all sources. Outputs are grouped by category for curriculum pretraining.
25
+
26
+ - **Total Kept:** 1,372,185 examples
27
+ - **Total Dropped >4096:** 1,208 examples
28
+ - **Total Duplicates Dropped:** ~181k+ (including shard/monolith duplicates)
29
+ - **Context:** ≤4096 tokens (cl100k_base proxy, hard ceiling)
30
+ - **Dedup:** SHA256 of normalized lowercased text, global across all datasets
31
+
32
+ ### Source Datasets (31)
33
+
34
+ All audited recursively including subfolders `/data`, `.jsonl` / `.json` / `.parquet` / `.jsonl.gz` with automatic gzip sniffing.
35
+
36
+ ```
37
+ 11-47/god_agent_grok4.4_cot_traces_20k
38
+ 11-47/claude_opus_mythos_5k
39
+ 11-47/claude_opus_4.8_max_thinking_5k_v2
40
+ 11-47/cluade_mythos_preview_5k_v2
41
+ 11-47/claude_opus_4.8_distill_5k
42
+ 11-47/god_level_python_dataset_25k
43
+ 11-47/Royal_Ghost_Coder_500k
44
+ 11-47/AI_Fiends_4_This_50k
45
+ WithinUsAI/fable_5_distillation_merged_cleaned_25k
46
+ WithinUsAI/MiniMax_M2.7_Distilled_5k
47
+ WithinUsAI/Llama_4_Maverick_Distilled_5k
48
+ WithinUsAI/Meta_Muse_Spark_Distilled_5k
49
+ WithinUsAI/kimi_2.6_Thinking_Distilled_5k
50
+ WithinUsAI/Qwen3.7_Max_Thinking_dataset_5K
51
+ WithinUsAI/GLM_5.1_Thinking_Distilled_5k
52
+ WithinUsAI/DeepSeek_V4_Flash_distilled_dataset_5k
53
+ WithinUsAI/gemini_3.5_flash_distilled_25k
54
+ WithinUsAI/grok_frontier_dataset_v3_100k
55
+ WithinUsAI/microsoft_copilot_distilled_25k
56
+ WithinUsAI/DEEPMIND_Alpha_Distilled
57
+ WithinUsAI/claude_mythos_distilled_25k
58
+ WithinUsAI/Gemini_3.2_Pro_Distilled
59
+ WithinUsAI/GPT_5.5_Distilled
60
+ WithinUsAI/Grok_4.4_Distilled
61
+ WithinUsAI/claude_Opus_4.7_Distilled
62
+ WithinUsAI/Opus4.7_thinking_max_distill_god_seed_25k
63
+ WithinUsAI/GOD_Coder_Complete_DataSet
64
+ WithinUsAI/Omega_Genesis_Coder_100k
65
+ WithinUsAI/Elite_GOD_Coder_100k
66
+ WithinUsAI/python_GOD_coder_100k
67
+ WithinUsAI/GOD_Coder_100k
68
+ ```
69
+
70
+ ### Build Stats (from your Kaggle run)
71
+
72
+ | Dataset | Kept | Dropped >4096 | Dropped Dupe |
73
+ |---|---|---:|---:|
74
+ | god_agent_grok4.4_cot_traces_20k | 11 | 0 | 19989 |
75
+ | claude_opus_mythos_5k | 24 | 0 | 4976 |
76
+ | claude_opus_4.8_max_thinking_5k_v2 | 4982 | 0 | 18 |
77
+ | cluade_mythos_preview_5k_v2 | 4986 | 0 | 14 |
78
+ | claude_opus_4.8_distill_5k | 5000 | 0 | 0 |
79
+ | god_level_python_dataset_25k | 276 | 0 | 24724 |
80
+ | Royal_Ghost_Coder_500k | 434124 | 0 | 67876 |
81
+ | AI_Fiends_4_This_50k | 50000 | 0 | 0 |
82
+ | fable_5_distillation_merged_cleaned_25k | 25146 | 573 | 0 |
83
+ | MiniMax_M2.7_Distilled_5k | 5000 | 0 | 0 |
84
+ | Llama_4_Maverick_Distilled_5k | 5000 | 0 | 0 |
85
+ | Meta_Muse_Spark_Distilled_5k | 545 | 0 | 4455 |
86
+ | kimi_2.6_Thinking_Distilled_5k | 5000 | 0 | 0 |
87
+ | Qwen3.7_Max_Thinking_dataset_5K | 5000 | 0 | 0 |
88
+ | GLM_5.1_Thinking_Distilled_5k | 4991 | 0 | 9 |
89
+ | DeepSeek_V4_Flash_distilled_dataset_5k | 5099 | 0 | 0 |
90
+ | gemini_3.5_flash_distilled_25k | 25000 | 0 | 0 |
91
+ | grok_frontier_dataset_v3_100k | 98938 | 0 | 1062 |
92
+ | microsoft_copilot_distilled_25k | 4596 | 0 | 20404 |
93
+ | DEEPMIND_Alpha_Distilled | 2297 | 0 | 1703 |
94
+ | claude_mythos_distilled_25k | 5315 | 0 | 19685 |
95
+ | Gemini_3.2_Pro_Distilled | 19896 | 7 | 0 |
96
+ | GPT_5.5_Distilled | 18013 | 184 | 0 |
97
+ | Grok_4.4_Distilled | 15386 | 256 | 0 |
98
+ | claude_Opus_4.7_Distilled | 28300 | 188 | 0 |
99
+ | Opus4.7_thinking_max_distill_god_seed_25k | 25000 | 0 | 0 |
100
+ | GOD_Coder_Complete_DataSet | 173860 | 0 | 1140 |
101
+ | Omega_Genesis_Coder_100k | 100100 | 0 | 0 |
102
+ | Elite_GOD_Coder_100k | 100100 | 0 | 0 |
103
+ | python_GOD_coder_100k | 100100 | 0 | 0 |
104
+ | GOD_Coder_100k | 100100 | 0 | 0 |
105
+
106
+ ### Category Organization
107
+
108
+ The builder classifies every kept example with this priority:
109
+
110
+ 1. **Tool_Calling** - contains `tool_calls`, `function_call`, or `<tool_call`
111
+ 2. **Think** - contains `<think>...</think>`
112
+ 3. **Thought** - contains `<thought>...</thought>` (distinct from Think)
113
+ 4. **Reasoning** - CoT markers without think tags
114
+ 5. **Code_Instruct** - name hint `coder/code/python`
115
+ 6. **Chat** - multi-turn `messages/conversations`
116
+ 7. **Instruct** - everything else
117
+
118
+ Output layout on Hub (add-only, never overwritten):
119
+
120
+ ```
121
+ Think/shard_00000.parquet
122
+ Think/shard_00001.parquet
123
+ ...
124
+ Thought/shard_00000.parquet
125
+ Reasoning/shard_00000.parquet
126
+ Code_Instruct/shard_00000.parquet
127
+ Code_Instruct/shard_00001.parquet
128
+ Instruct/shard_00000.parquet
129
+ Chat/shard_00000.parquet
130
+ Tool_Calling/shard_00000.parquet
131
+ ```
132
+
133
+ Each shard index is computed from the live repo listing, so re-running the builder never overwrites.
134
+
135
+ ### Schema
136
+
137
+ Each parquet row:
138
+
139
+ - `text: string` - flattened training text, ≤4096 tokens, ready for pretraining
140
+ - `n_tokens: int` - token count via cl100k_base
141
+ - `category: string` - one of [Think, Thought, Reasoning, Code_Instruct, Tool_Calling, Chat, Instruct]
142
+ - `source_dataset: string` - original HF repo id
143
+ - `raw: string` - original record JSON (truncated to 20k chars)
144
+
145
+ ### How it was built (Kaggle-safe)
146
+
147
+ - Batch size 5 datasets at a time, deletes raw download after each dataset
148
+ - Metadata files (`manifest.json`, `dataset_infos.json`) skipped
149
+ - Monolith `train.jsonl` that duplicates shard family `train-00000-of-...jsonl` skipped
150
+ - Gzip magic byte sniffing (`0x1f8b`) for files misnamed as `.jsonl`
151
+ - Global dedup set rebuilt from existing Hub shards on every resume (Kaggle `/kaggle/working` is not durable)
152
+ - Uses `HF_TOKEN` from Kaggle Secrets label `ITLL`
153
+
154
+ ### Usage
155
+
156
+ ```python
157
+ from datasets import load_dataset
158
+
159
+ # load everything
160
+ ds = load_dataset("ITLL/Organized_PreTrain_WIUAI")
161
+
162
+ # load only code
163
+ code = load_dataset("ITLL/Organized_PreTrain_WIUAI", data_dir="Code_Instruct")
164
+
165
+ # load Think + Reasoning for curriculum
166
+ think = load_dataset("ITLL/Organized_PreTrain_WIUAI", data_dir="Think")
167
+ reason = load_dataset("ITLL/Organized_PreTrain_WIUAI", data_dir="Reasoning")
168
+
169
+ # pretrain streaming
170
+ from datasets import load_dataset
171
+ stream = load_dataset("ITLL/Organized_PreTrain_WIUAI", streaming=True, split="train")
172
+ for ex in stream:
173
+ print(ex["category"], ex["n_tokens"])
174
+ ```
175
+
176
+ ### Intended Use
177
+
178
+ 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.
179
+
180
+ ### Builder Code
181
+
182
+ Full Kaggle cell is stored in `builder.py` in this repo if you want to reproduce. Requires `huggingface_hub`, `pandas`, `pyarrow`, `tiktoken`.
183
+
184
+ ### License
185
+
186
+ MIT - inherits from source distillations. Please check original dataset licenses for commercial use. This is a cleaned aggregation for research.