Dataset Viewer
Auto-converted to Parquet Duplicate
The dataset viewer is not available for this split.
Parquet error: Scan size limit exceeded: attempted to read 520842613 bytes, limit is 300000000 bytes Make sure that 1. the Parquet files contain a page index to enable random access without loading entire row groups2. otherwise use smaller row-group sizes when serializing the Parquet files
Error code:   TooBigContentError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

ForoCoches long-post crawl 2026

Every ForoCoches thread ID from 1 to 10,841,474 was requested once (October 2026, signed-in account, first page of each thread only). From each first page we kept the long posts (≥2,000 characters of the author's own text, by anyone), the opening post's other posts in those threads, and short opening posts that give a long post its context. Everything else on the page was dropped at collection time.

The source for forocoches-lm v3, a Gemma 3 1B model that writes forum threads. The curated, deduplicated training pool is forocoches-threads-v3.

threads 345,039
posts 1,424,210
long posts (≥2,000 own chars) 482,818
years (opening post) 2003–2026

Configs

threads (default): one document per thread in the forum format used for training:

TÍTULO: <title>

#1 OP: <opening post>

#2 forero_a: <long reply>

#3 OP: <the author continues>

Posts are numbered in reading order among the kept posts, not by forum position. OP follow-ups shorter than 280 characters of own text are left out of the document (they mostly answer replies that were not kept), but they stay in posts.

field
thread_id, page_url, collected_at provenance
title, year, tags, has_op has_op is false when only replies were long (the opener was not kept)
text the document above
post_ids source posts in the document
tipo, calidad, label_model thread type and quality from Claude Haiku (title + first 1,500 opener characters + three short replies). Empty for the 19,981 threads that were already in the historical archive and for 149 the classifier declined.

posts: every kept post: thread_id, post_id, position (forum position on page 1), author (per-thread pseudonym), retention (long, op_followup, op_context), posted_at (as shown by the forum), own_chars, text, title.

Cleaning and pseudonyms

HTML → plain text: quotes become [cita=forero_x]…[/cita] (trimmed to 200 characters), images/videos/links/emails/phones become [img], [video …], [enlace dominio], [email], [teléfono]. Authors are OP, forero_a, forero_b… per thread. Usernames of five or more characters are replaced wherever they appear in the text; @mentions and plain-text quote headers (Cita de …, Escrito originalmente por …) of users who are not in the thread become usuario. Usernames shorter than five characters are left, because they collide with ordinary words; titles are not rewritten.

Caveats

  • First pages only: long stories that continue on later pages are cut.
  • Labels are automated and unaudited for this crawl; 92k threads were labelled noticia (pasted articles) and 15k compraventa.
  • Forum text includes rude, sexist and otherwise offensive content, kept as written.
Downloads last month
73