umi-robots

Part of umi, an open web crawl published as Parquet. Before you use any of this, read the exclusion list at open-index/umi-meta and filter the rows it names. Published files are never rewritten, so the exclusion list is how a takedown reaches you, and applying it is a condition of using the data rather than a suggestion.

One row per robots.txt fetch: the host, when we asked, what the origin answered, the raw text if it served one, and the summary our parser read out of it. It is a record of what sites told crawlers, kept over time, and as far as we know nobody else publishes it.

Columns

The schema is doc 10.5's, mapped one to one into Parquet with nothing renamed and nothing flattened.

column	type	null	what it is
host	string	no	The host, with no scheme on it.
fetched_at_ms	uint64	no	When we asked, Unix milliseconds.
status	uint16	no	What the origin answered. 404 is the common case and it means the site has no file rather than that we failed to reach it. Zero means we never got a response at all, which is the one case where the rules below are ours and not the site's.
body	string	yes	The raw file, exactly as served. Null unless the status was a 2xx, because a 404 body is somebody's error page and not a robots.txt.
groups	uint32	no	How many User-agent groups the file has, ours and everybody else's. Next to rules it says whether a site wrote rules for us or we are reading its wildcard group.
rules	uint32	no	How many rules were in the group that applied to us.
crawl_delay_ms	uint32	yes	The Crawl-delay that applied to us, in milliseconds. Null when the file did not set one.
allows_us	uint8	no	Whether the root path was allowed to umi, 1 or 0. A host that could not be reached is 0, because an unreadable robots.txt disallows rather than allows.
sitemaps	list of string	no	Every Sitemap line in the file, in order.
content_usage	string	yes	The AIPREF Content-Usage value the file declared, verbatim.

How this was made

Extractor umi/0.0.1, canonicalisation canon/1, against the spec at docs/spec. Rows are written in fetch completion order, so the fetched_at_ms statistics on a row group are the ones worth pushing a filter down to. Files are around 128 MB and a day folder holds one day of them.

umi crawls to build a public, openly licensed web index, and that is the whole declaration. We do not train models on the corpus, we do not run an agent that browses for a user, and we do not resell access. The crawler identifies itself as umi and its bot page is https://umi.dev/bot. The corpus is open, so what anyone else does with it is out of our hands, which is a fact worth stating plainly rather than eliding.

Robots and Content-Usage

Every page in the corpus was fetched under RFC 9309: the host's robots.txt was read first, a Disallow that matched was obeyed, and a robots.txt we could not read at all disallowed the host rather than allowing it. On the pages schema robots_checked_ms says when the file behind that decision was last read.

content_usage carries the site's AIPREF Content-Usage preference, from the response header or the robots.txt, verbatim and unparsed. We record it and we do not act on it, because it is the publisher's statement to whoever reads the corpus and not ours to interpret on their behalf. If you are training on this data, that column is the one to read.

A robots snapshot is only good for a day. RFC 9309 caps a cached robots.txt at 24 hours and umi honours that, so the robots corpus is a record of what a host said when we asked, not a standing permission you can crawl against today. Use it to plan, to see which hosts have rules at all, to pick up Crawl-delay and Sitemap before you queue anything, and then ask the origin yourself.

What a zero status means

Roughly two rows in five carry status = 0, and it does not mean the host has no robots.txt. A host with no file answers 404 and gets a row saying 404. A zero means the request came back with nothing at all: the name did not resolve, or it resolved and the connection was refused, reset, or timed out.

That matters because the two cases are very different and the column does not tell them apart. On the deep tail of the domain ranking most of the silence is names that no longer exist, and sampled against the ranking, 88 percent of dead apexes at rank two million and 97 percent at rank five and a half million were NXDOMAIN. Further down it inverts: on ranks past eighteen million, nearly half the hosts we ask are names that resolve and then serve nothing.

We ask twice. A host that answers nothing on the first request is asked once more when its name resolves, and roughly three in ten of those second asks come back with something, so close to a third of what would otherwise be silence is a real answer. Rows fetched before 2026-09-02 were one attempt only and carry more transient zeros than later rows do. fetched_at_ms is how you tell them apart.

So treat a zero as "we did not get an answer" and not as a statement by the site. In particular allows_us is 0 on those rows because RFC 9309 says an unreadable robots.txt disallows rather than allows, which is our rule applied to our own failure. It is not the host refusing you.

Reading it

DuckDB over the files in place, without downloading the repository:

CREATE SECRET (TYPE huggingface, PROVIDER credential_chain);

SELECT host, crawl_delay_ms, sitemaps
FROM read_parquet('hf://datasets/open-index/umi-robots/data/**/*.parquet')
WHERE status = 200 AND crawl_delay_ms IS NOT NULL
LIMIT 20;

The files are around 128 MB and a day folder holds one day of them, so a query that names a day reads a day. Rows are in fetch completion order, which means fetched_at_ms is the column whose row group statistics are worth pushing a filter down to, and host is not: a run walks the domain ranking rather than the alphabet, so hosts are spread across every file.

One host can appear more than once, though as the corpus stands it rarely does: counted on 2026-09-06 there were 26,987,785 rows over 26,863,085 distinct hosts, so 0.46 percent of the rows are a second answer for a host already in the corpus. Four days earlier it was 23.1 percent. That was not history. Every one of those rows was fetched inside a single day by two runs whose slices of the domain ranking overlapped, because nothing in the crawler stopped a run asking a host an earlier run had already answered. Something does now, the files carrying most of the repeats were taken back out, and the hosts whose only answer was in one of them were asked again. The removal is recorded under retractions/ in the meta repository, with every removed file named and the digest it had.

What is left is small enough to ignore for most work and it is still there, so take the newest fetched_at_ms per host if what you want is one current answer each:

SELECT host, status, crawl_delay_ms, allows_us
FROM read_parquet('hf://datasets/open-index/umi-robots/data/**/*.parquet')
QUALIFY row_number() OVER (PARTITION BY host ORDER BY fetched_at_ms DESC) = 1;

That reads every file, so narrow the path to the days you want if you do not need all of it.

Corrections

We never rewrite a published file, because that breaks every checksum anyone recorded and makes old work unreproducible. Corrections go on an append only exclusion list in open-index/umi-meta under blocks/, naming a repository and a file and either a row predicate or a set of url_key values, with a reason and a date. Apply it as a filter when you read.

Every day folder has a manifest under _manifest/ listing each file with its digests, and a detached Ed25519 signature next to it. The signing keys are in the same meta repository, so a manifest can be checked without asking us for anything.

A file being removed is meant to be rare and it is not impossible. When it happens the deletion and the rewritten day manifests go in one commit, the manifest chain is relinked to the end of the repository, and a record naming every removed file with the digest it had goes to the same meta repository under retractions/ first. So if a digest you recorded stops matching, look there before concluding you were served something altered. Nothing we can do repairs the digest itself.

Licence

CC0. This is a record umi produced rather than content it collected, so there is nothing here anybody else holds a right in.

Takedown

Mail tamnd87@gmail.com with the URLs or the host. A request is honoured by adding the rows to the exclusion list, which takes minutes rather than a release cycle, and by adding the host to the crawler's block list so it is not fetched again. You do not need to explain why.