| --- |
| language: [en] |
| size_categories: [n<1K] |
| tags: [pretraining, essential-web, diffusion-lm] |
| --- |
| |
| # essential-web-subsample |
|
|
| A small, category-balanced subsample of [Essential-Web v1.0](https://huggingface.co/datasets/EssentialAI/essential-web-v1.0) for studying how generation behavior (e.g. diffusion-LM parallelism) varies with data type. |
|
|
| **900 documents = 100 per category** across 9 categories, mapped from Essential-Web's `document_type_v1.primary.label`: |
|
|
| | category | Essential-Web document_type_v1 | |
| |---|---| |
| | code | Code/Software | |
| | papers | Academic/Research | |
| | encyclopedic | Reference/Encyclopedic/Educational | |
| | legal | Legal/Regulatory | |
| | literary | Literary/Creative | |
| | news | News/Editorial | |
| | social | Social/Forum | |
| | reviews | Reviews/Critiques | |
| | political | Government/Political | |
|
|
| Each document is **word-level truncated so its token count is in [1024, 2048] under BOTH the LLaDA-8B-Base and Dream-v0-Base-7B tokenizers** — so the identical text is usable by both models (no tokenizer-specific truncation). |
|
|
| **Fields:** `text`, `category`, `n_tokens_llada`, `n_tokens_dream`, `id`, `url`. |
|
|
| **Source:** documents drawn from the `CC-MAIN-2024-38` crawl of Essential-Web v1.0. Note: Essential-Web is web-derived, so a category like `code` means web pages classified as code/software content (dev blogs, docs, forum posts), not raw source files. Inherits Essential-Web's license/terms. |
|
|
