File size: 6,702 Bytes
bedf153 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 | ---
license: mit
task_categories:
- text-generation
language:
- en
tags:
- pretraining
- reasoning
- chain-of-thought
- code
- instruct
- tool-calling
- chat
- think
size_categories:
- 1M<n<10M
---
# ITLL/Organized_PreTrain_WIUAI
**A clean, deduped, 4096-context pretrain-ready merge of 31 distillation datasets from 11-47 and WithinUsAI.**
All examples >4096 tokens were **100% trashed, never truncated**. Duplicates were removed globally across all sources. Outputs are grouped by category for curriculum pretraining.
- **Total Kept:** 1,372,185 examples
- **Total Dropped >4096:** 1,208 examples
- **Total Duplicates Dropped:** ~181k+ (including shard/monolith duplicates)
- **Context:** ≤4096 tokens (cl100k_base proxy, hard ceiling)
- **Dedup:** SHA256 of normalized lowercased text, global across all datasets
### Source Datasets (31)
All audited recursively including subfolders `/data`, `.jsonl` / `.json` / `.parquet` / `.jsonl.gz` with automatic gzip sniffing.
```
11-47/god_agent_grok4.4_cot_traces_20k
11-47/claude_opus_mythos_5k
11-47/claude_opus_4.8_max_thinking_5k_v2
11-47/cluade_mythos_preview_5k_v2
11-47/claude_opus_4.8_distill_5k
11-47/god_level_python_dataset_25k
11-47/Royal_Ghost_Coder_500k
11-47/AI_Fiends_4_This_50k
WithinUsAI/fable_5_distillation_merged_cleaned_25k
WithinUsAI/MiniMax_M2.7_Distilled_5k
WithinUsAI/Llama_4_Maverick_Distilled_5k
WithinUsAI/Meta_Muse_Spark_Distilled_5k
WithinUsAI/kimi_2.6_Thinking_Distilled_5k
WithinUsAI/Qwen3.7_Max_Thinking_dataset_5K
WithinUsAI/GLM_5.1_Thinking_Distilled_5k
WithinUsAI/DeepSeek_V4_Flash_distilled_dataset_5k
WithinUsAI/gemini_3.5_flash_distilled_25k
WithinUsAI/grok_frontier_dataset_v3_100k
WithinUsAI/microsoft_copilot_distilled_25k
WithinUsAI/DEEPMIND_Alpha_Distilled
WithinUsAI/claude_mythos_distilled_25k
WithinUsAI/Gemini_3.2_Pro_Distilled
WithinUsAI/GPT_5.5_Distilled
WithinUsAI/Grok_4.4_Distilled
WithinUsAI/claude_Opus_4.7_Distilled
WithinUsAI/Opus4.7_thinking_max_distill_god_seed_25k
WithinUsAI/GOD_Coder_Complete_DataSet
WithinUsAI/Omega_Genesis_Coder_100k
WithinUsAI/Elite_GOD_Coder_100k
WithinUsAI/python_GOD_coder_100k
WithinUsAI/GOD_Coder_100k
```
### Build Stats (from your Kaggle run)
| Dataset | Kept | Dropped >4096 | Dropped Dupe |
|---|---|---:|---:|
| god_agent_grok4.4_cot_traces_20k | 11 | 0 | 19989 |
| claude_opus_mythos_5k | 24 | 0 | 4976 |
| claude_opus_4.8_max_thinking_5k_v2 | 4982 | 0 | 18 |
| cluade_mythos_preview_5k_v2 | 4986 | 0 | 14 |
| claude_opus_4.8_distill_5k | 5000 | 0 | 0 |
| god_level_python_dataset_25k | 276 | 0 | 24724 |
| Royal_Ghost_Coder_500k | 434124 | 0 | 67876 |
| AI_Fiends_4_This_50k | 50000 | 0 | 0 |
| fable_5_distillation_merged_cleaned_25k | 25146 | 573 | 0 |
| MiniMax_M2.7_Distilled_5k | 5000 | 0 | 0 |
| Llama_4_Maverick_Distilled_5k | 5000 | 0 | 0 |
| Meta_Muse_Spark_Distilled_5k | 545 | 0 | 4455 |
| kimi_2.6_Thinking_Distilled_5k | 5000 | 0 | 0 |
| Qwen3.7_Max_Thinking_dataset_5K | 5000 | 0 | 0 |
| GLM_5.1_Thinking_Distilled_5k | 4991 | 0 | 9 |
| DeepSeek_V4_Flash_distilled_dataset_5k | 5099 | 0 | 0 |
| gemini_3.5_flash_distilled_25k | 25000 | 0 | 0 |
| grok_frontier_dataset_v3_100k | 98938 | 0 | 1062 |
| microsoft_copilot_distilled_25k | 4596 | 0 | 20404 |
| DEEPMIND_Alpha_Distilled | 2297 | 0 | 1703 |
| claude_mythos_distilled_25k | 5315 | 0 | 19685 |
| Gemini_3.2_Pro_Distilled | 19896 | 7 | 0 |
| GPT_5.5_Distilled | 18013 | 184 | 0 |
| Grok_4.4_Distilled | 15386 | 256 | 0 |
| claude_Opus_4.7_Distilled | 28300 | 188 | 0 |
| Opus4.7_thinking_max_distill_god_seed_25k | 25000 | 0 | 0 |
| GOD_Coder_Complete_DataSet | 173860 | 0 | 1140 |
| Omega_Genesis_Coder_100k | 100100 | 0 | 0 |
| Elite_GOD_Coder_100k | 100100 | 0 | 0 |
| python_GOD_coder_100k | 100100 | 0 | 0 |
| GOD_Coder_100k | 100100 | 0 | 0 |
### Category Organization
The builder classifies every kept example with this priority:
1. **Tool_Calling** - contains `tool_calls`, `function_call`, or `<tool_call`
2. **Think** - contains `<think>...</think>`
3. **Thought** - contains `<thought>...</thought>` (distinct from Think)
4. **Reasoning** - CoT markers without think tags
5. **Code_Instruct** - name hint `coder/code/python`
6. **Chat** - multi-turn `messages/conversations`
7. **Instruct** - everything else
Output layout on Hub (add-only, never overwritten):
```
Think/shard_00000.parquet
Think/shard_00001.parquet
...
Thought/shard_00000.parquet
Reasoning/shard_00000.parquet
Code_Instruct/shard_00000.parquet
Code_Instruct/shard_00001.parquet
Instruct/shard_00000.parquet
Chat/shard_00000.parquet
Tool_Calling/shard_00000.parquet
```
Each shard index is computed from the live repo listing, so re-running the builder never overwrites.
### Schema
Each parquet row:
- `text: string` - flattened training text, ≤4096 tokens, ready for pretraining
- `n_tokens: int` - token count via cl100k_base
- `category: string` - one of [Think, Thought, Reasoning, Code_Instruct, Tool_Calling, Chat, Instruct]
- `source_dataset: string` - original HF repo id
- `raw: string` - original record JSON (truncated to 20k chars)
### How it was built (Kaggle-safe)
- Batch size 5 datasets at a time, deletes raw download after each dataset
- Metadata files (`manifest.json`, `dataset_infos.json`) skipped
- Monolith `train.jsonl` that duplicates shard family `train-00000-of-...jsonl` skipped
- Gzip magic byte sniffing (`0x1f8b`) for files misnamed as `.jsonl`
- Global dedup set rebuilt from existing Hub shards on every resume (Kaggle `/kaggle/working` is not durable)
- Uses `HF_TOKEN` from Kaggle Secrets label `ITLL`
### Usage
```python
from datasets import load_dataset
# load everything
ds = load_dataset("ITLL/Organized_PreTrain_WIUAI")
# load only code
code = load_dataset("ITLL/Organized_PreTrain_WIUAI", data_dir="Code_Instruct")
# load Think + Reasoning for curriculum
think = load_dataset("ITLL/Organized_PreTrain_WIUAI", data_dir="Think")
reason = load_dataset("ITLL/Organized_PreTrain_WIUAI", data_dir="Reasoning")
# pretrain streaming
from datasets import load_dataset
stream = load_dataset("ITLL/Organized_PreTrain_WIUAI", streaming=True, split="train")
for ex in stream:
print(ex["category"], ex["n_tokens"])
```
### Intended Use
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.
### Builder Code
Full Kaggle cell is stored in `builder.py` in this repo if you want to reproduce. Requires `huggingface_hub`, `pandas`, `pyarrow`, `tiktoken`.
### License
MIT - inherits from source distillations. Please check original dataset licenses for commercial use. This is a cleaned aggregation for research.
|