halo-hil

Web text in hil, re-filtered by language and prepared for pretraining.

What changed, and why it had to

The earlier version of this dataset was labelled hil by the crawler's own language detection, and that label was never verified. An audit on 2026-09-22 found that most of it was not hil: over a random sample of 1,499 sentences, GlotLID v3 called 44 % English, 22 % Filipino/Tagalog and only 12 % Hiligaynon — much of the corpus was Tagalog news copy and site boilerplate.

That mattered beyond this dataset. A language identifier trained on these labels learned that Tagalog news is Hiligaynon, and then rejected 5,747 genuinely Filipino pages as Hiligaynon in a downstream scrape.

This revision re-gates every document by language:

step	what it does
clean	the repo's shared boilerplate cleaner (halolib.cleaner)
split	documents into sentences
identify	every sentence, with halo-lid deciding and GlotLID v3 as a second opinion, giving a character-weighted share per language
decide	keep a document whole when ≥50% of it is hil; salvage the hil sentences when the document is ≥15% but mixed; drop the rest
deduplicate	md5 exact, then MinHash LSH near-duplicates at Jaccard 0.75

Salvage exists because bilingual Philippine news sites publish the same article in English and in the local language on one page. Dropping the page loses real hil; keeping it poisons the corpus. The extraction column records which documents were salvaged so you can exclude them.

Size

	documents
before	9,860	unverified label
after	500	5.1% retained, 176,494 words
├ kept whole	397	≥50% hil
└ salvaged	103	hil sentences taken from mixed pages

Dropped: 8,464 off-language, 1 too short after filtering, 895 duplicates.

Splits: 450 train / 50 test.

Columns

column	meaning
text	cleaned text, language-filtered — use this for pretraining
language	hil, now verified rather than assumed
extraction	whole or salvaged
lid_target_share	share of the original document's characters identified as hil
lid_share	the full per-language share, as JSON
lid_glotlid_doc	GlotLID's verdict on the whole original document
url, date, dump, source	provenance
word_count, token_count, content_hash	FineWeb-compatible fields

Honest limitations

• The filter is only as good as the identifier. halo-lid scores 0.853 on Hiligaynon against human-labelled read prompts, so a few percent of these sentences are probably still misfiled, and some genuine hil was certainly thrown away.
• Salvaged documents are sentence sequences from a page, not continuous prose. They are fine for language modelling and poor for anything needing document structure.
• This is web text: the domain skews to news and blogs, and it has had no quality scoring beyond length and language.
• The previous, unfiltered revision is still in this repository's git history at 6c35dc63f992 if you need it for reproducibility.

Produced by scripts/refilter_corpus.py in the halohalo pipeline. Report: refilter_report.json.