README.md

metadata

pretty_name: Wikipethia
license: other
license_name: mixed-per-source
license_link: https://github.com/JossDuff/wikipethia#licensing
language:
  - en
tags:
  - ethereum
  - blockchain
  - research
  - specifications
  - forum
size_categories:
  - 10K<n<100K
task_categories:
  - text-retrieval
  - question-answering

Wikipethia

A curated corpus of Ethereum research and standards: nine years of ethresear.ch and Ethereum Magicians discussion, the EIPs and ERCs, the consensus- and execution-layer specifications, the engine API, AllCoreDevs meeting notes, Vitalik Buterin's writing, and the Ethereum Foundation blog. One row per document, with the author, date, URL, and license carried on every row.

Built by wikipethia, an MCP server that gives LLM clients hybrid search over this corpus. This dataset is the corpus itself, published for anyone who wants to build on it — fine-tuning, retrieval, or analysis.

Snapshot date: 2026-08-24. Everything here reflects the sources as of that date; see Removals and corrections below.

Loading

from datasets import load_dataset

ds = load_dataset("JossDuff/wikipethia", split="train")

The corpus mixes licenses (see the table below). The license column makes filtering trivial — for example, to keep only public-domain rows:

cc0 = ds.filter(lambda r: r["license"] == "CC0-1.0")

or everything available for commercial use (all sources except the two forums):

commercial = ds.filter(lambda r: r["license"] != "CC-BY-NC-SA-3.0")

Schema

Column	Type	Description
id	string	Stable document id, source/… (e.g. ethresearch/post/39462, eips/eip-4844)
source	string	Which source the document came from (see table below)
license	string	The source's license, denormalized onto every row
url	string	Canonical URL of the document
title	string	Document title; forum posts carry their topic's title
author	string, nullable	Author as the source states it (forum username, EIP author line); null where the source has none (2,571 rows, mostly spec files)
published	string, nullable	ISO 8601 publication timestamp; null for 1 row
content	string	The document body, as markdown (see preprocessing notes)
meta	string	Source-specific metadata as a JSON object (e.g. topic_id, post_number, tags, EIP status); keys vary by source, {} where none

Sources and licensing

Every license below comes from the publisher's own statement, not inferred. The dataset as a whole is therefore usable for non-commercial purposes with attribution; the non-forum subset (~6%) is usable commercially. Attribution comes for free: every row carries its author, date, and URL.

source	What it is	Rows	License
ethmagicians	Ethereum Magicians forum posts	33,514	CC BY-NC-SA 3.0
ethresearch	ethresear.ch forum posts	20,217	CC BY-NC-SA 3.0
executionspecs	ethereum/execution-specs — EELS, the executable execution-layer spec	1,046	CC0 1.0
efblog	Ethereum Foundation blog	634	CC BY 4.0
ercs	ethereum/ERCs — application-layer standards	612	CC0 1.0
pm	ethereum/pm — AllCoreDevs agendas and notes	612	CC BY-SA 3.0
eips	ethereum/EIPs — core protocol proposals	585	CC0 1.0
vitalik	vitalik.eth.limo — Vitalik Buterin's writing	175	WTFPL
consensusspecs	ethereum/consensus-specs — the beacon chain spec, per fork	93	CC0 1.0
executionapis	ethereum/execution-apis — the engine API	11	CC0 1.0

On training: whether a model trained on share-alike or non-commercial text inherits those terms is an unsettled legal question. This card states what each source's license is and gives you the column to filter on; the judgement is yours.

Provenance and preprocessing

Content is fetched from each source's own API or repository (Discourse API for the forums, GitHub tarballs for the spec repos, feeds for the blogs) at one request per second. What was done to it, so you inherit it knowingly:

• Forum posts are the raw markdown, not rendered HTML. MathJax ($$...$$) survives intact — ethresear.ch is heavy on it.
• One row per post, not per thread. A thread's replies share the topic title; meta.topic_id and meta.post_number reconstruct threads.
• Quote blocks ([quote="…"]) were stripped from forum posts at ingest — they duplicate text across rows.
• Spec repos are one row per file, markdown and (for execution-specs) Python. The consensus and execution specs repeat per fork directory, so near-identical fork copies of the same document are present by construction.
• Some forum posts are stubs pointing at HackMD, arXiv, or a blog — the substance lives behind the link, not in content.
• EIPs that moved to the ERCs repository appear under ercs, not eips — resolve a proposal number by searching both sources.
• Deleted posts and upstream edits are picked up at sync time, so this snapshot reflects the sources as of the snapshot date — but not after it.

Things to know before you train on this

• Ethereum research goes stale in specific ways. A 2019 sharding post can flatly contradict a 2024 one, and both are in here. The published column is not optional metadata — anything reasoning over this corpus needs it to handle supersession.
• The two forums are 93% of the rows. Sampling uniformly means training almost entirely on forum discussion; use source to rebalance if you want spec-shaped text represented.
• Documents range from a one-character forum reply to a 140k-character EIP. Length varies by orders of magnitude; chunk to your own needs — no chunking has been applied here.
• Forum usernames are public but personal. Authors posted publicly under these names, and the forums' licenses require attribution — but be thoughtful about uses that profile individuals rather than the research.

Removals and corrections

Upstream, a forum user can delete a post or anonymize their account; this snapshot cannot see anything that happened after its date. New snapshots are published from fresh syncs, which pick up deletions. If something here should be removed, open an issue at JossDuff/wikipethia and it will be dropped from the next snapshot.

What this dataset is not

It is the documents only. Wikipethia's search index, chunking, and embeddings are not included: chunking decisions belong to the consumer, and embedding vectors are only meaningful to the exact model that produced them. To get the ready-made hybrid-search MCP server over this corpus, see the wikipethia repository.