LocalDoc/community_oscar_azerbaijani_scored

Azerbaijani Web Corpus with Quality Scores

This dataset is the full Azerbaijani web corpus LocalDoc/community_oscar_azerbaijani with a continuous quality score attached to every document. It is intended as the filtering layer for building a clean Azerbaijani pretraining corpus: each document carries a score that lets you keep, clean, or drop it according to your own thresholds.

What was done

Every document in the source corpus was scored by the model LocalDoc/azerbaijani-text-quality-classifier, a regression model (mmBERT-base backbone) that rates how much genuine, well-formed Azerbaijani prose a document contains.

• Source corpus: LocalDoc/community_oscar_azerbaijani (~24M documents, OSCAR-derived Common Crawl)
• Scoring model: LocalDoc/azerbaijani-text-quality-classifier
• Compute: scoring took ~4 days on a single RTX 6000 Ada
• Max sequence length at scoring: 4096 tokens (longer documents truncated)

Fields

• index — identifier of the form <source_shard>:<row> (stable, unique across the dataset)
• score — continuous quality score, regression output (roughly -2.5 .. 3.75)
• content — the document text

The score

The score is a continuous regression value, NOT a discrete 0/1/2/3 class. It roughly follows the training scale where higher means cleaner text:

• ~3 — clean, coherent Azerbaijani prose
• ~2 — substantial good prose mixed with junk (menus, footers, ads)
• ~1 — mostly junk, little recoverable prose
• ~0 and below — pure junk: navigation, spam, machine translation, non-Azerbaijani text

Values can fall slightly outside [0, 3] because the regression head is not clipped.

Score distribution (full corpus)

• documents: ~24,072,723
• mean: 1.83, median: 1.86
• share with score >= 1.0: ~87%
• share with score >= 2.0: ~39%
• share with score >= 2.5: ~22%
• share with score >= 3.0: ~4%

The distribution is roughly bimodal, with a mode near 1.8-2.0 (mixed documents) and a separate mode near 2.8-3.0 (clean documents), and a dip around 2.3-2.5.

Suggested use

Pick thresholds for a three-way routing rather than a single cutoff:

• drop documents below a low threshold (pure junk)
• clean mid-range documents with line-level rules (they contain good prose wrapped in site boilerplate — see limitations)
• keep high-scoring documents as-is

The score lets you set and re-set thresholds without re-running the expensive scoring pass.

Limitations

• Scores are model-generated, not human-verified. The scoring model was trained on LLM-generated labels, and its agreement with human judgement has not been measured on a human-annotated test set.
• Clean prose wrapped in heavy boilerplate is under-scored. A clean encyclopedic article surrounded by navigation/footers can receive a low score (e.g. a Wikipedia article scored ~1.2). Such documents should be routed to line-level cleaning, not dropped — do not use a naive single threshold like ">= 2", or you will discard good articles.
• Scores are noisier in the mid-range (roughly 0.5-1.5): near-identical documents can differ by ~0.5. Treat mid-range scores as approximate.
• The corpus is raw web crawl and may contain personal data; no PII removal was applied.

Source and attribution

Texts originate from LocalDoc/community_oscar_azerbaijani. This dataset adds only the score column and the index identifier.