---
pretty_name: Commonplace Knowledge — Simple English Wikipedia
language:
- en
license: other
license_name: component-specific-apache-2.0-and-cc-by-sa
license_link: https://huggingface.co/datasets/DevParker/Commonplace-Knowledge/blob/main/LICENSES.md
task_categories:
- text-retrieval
tags:
- wikipedia
- embeddings
- local-research
- source-linked-memory
size_categories:
- 100K<n<1M
configs:
- config_name: articles
  data_files:
  - split: train
    path: source/simplewiki.parquet
- config_name: passages
  data_files:
  - split: train
    path: index/chunks.parquet
---

# Commonplace Knowledge

A reproducible starting knowledge pack for Commonplace, a local research and
memory tool. Published as a Hugging Face **dataset** repository at
[DevParker/Commonplace-Knowledge](https://huggingface.co/datasets/DevParker/Commonplace-Knowledge).
It includes the pinned encoder as a component; it is not a newly trained model.

The payload contains **241,787 Simple English Wikipedia articles**, **344,558
passages**, **5,772 narrow extracted assertions**, a vector index, and a clean
writable-memory seed with no custom additions. No personal notes, saved model
thoughts, research transcripts, generation-model weights or credentials are included.

## Contents

| Folder | Contents |
| --- | --- |
| `source/` | Original pinned Simple English snapshot Parquet |
| `index/` | Source-offset passages, normalized 512-dimensional FAISS index, metadata |
| `catalog/` | Immutable lexical/entity/assertion SQLite catalog and active manifest |
| `seed/` | Empty writable-memory schema; no custom additions |
| `encoder/` | Pinned Google EmbeddingGemma 2 model, tokenizer and configuration |
| `provenance/` | Original dataset card, identities, chunking and seed audit |
| `licenses/` | Applicable license texts |

`manifest.json` lists SHA-256 and byte size for every payload file. All payloads
are real local files, with no links into another project. Large files have Git LFS
attributes. The manifest records the exact total; the pack is approximately 3.3 GiB.

## Pinned sources

* Dataset: `wikimedia/wikipedia`, snapshot `20231101.simple`, revision
  `b04c8d1ceb2f5cd4588862100d08de323dccfbaa`.
* Encoder: `google/embeddinggemma-2`, revision
  `914f7f89142e33e77833254d9c9b90c3cef7303b`.

The index uses exact inner-product search over normalized FP32 vectors.
The metadata's `storage_dtype: float16` describes the original input shards,
not the internal FAISS index representation. Redundant input shards are excluded.

## Use

Install the independent [Commonplace MCP/API](https://github.com/BANTAM-ADMIN/commonplace).
It does not include or require Bantam Factory or a generation model. From the
software checkout, after installing its Python and Node dependencies:

```bash
.venv/bin/python scripts/download_pack.py --dest ../Commonplace-Knowledge
.venv/bin/python scripts/prepare_store.py --pack ../Commonplace-Knowledge
node scripts/commonplace.mjs serve
```

The software's release descriptor pins an exact dataset commit and pack identity.
No Hugging Face sign-in is required to download this public pack. See the
[installation and MCP connection guide](https://github.com/BANTAM-ADMIN/commonplace/blob/main/docs/QUICKSTART.md).

Setup verifies checksums and copies the public seed into a separate writable
local store. Subsequent research changes that store, not this public pack.
The encoder runs on CPU by default. A local generation model is needed only for
the optional full researcher, not for direct memory access.

The dataset card declares two separate Parquet configurations, `articles` and
`passages`. If using the Hugging Face `datasets` library after publication, choose
one explicitly; they have different schemas. Commonplace itself downloads the
whole pinned pack and uses its local SQLite/FAISS files, without that library.

## Coverage and limitations

This is Simple English Wikipedia from November 2023. It does not contain the
complete English Wikipedia and is not a current world-knowledge database.
Sparse articles, missing topics, ambiguous entities and
extraction errors remain. Similarity scores do not establish truth. A source
citation does not prove that a model correctly interpreted the source.

Wikipedia's editorial and geographic coverage is uneven. Short, simplified
articles can omit important qualifications, and the source may contain errors,
outdated information or sensitive material. This release does not independently
fact-check every article or extracted assertion.

The seed contains no model interpretations; these accrue only in a user's local
writable memory. No promise of automatic improvement or universal answer-quality
gains is made. See the software repository's validation and design documents.

Attribution, transformations and component licenses: [LICENSES.md](LICENSES.md).