license: cc-by-4.0
language:

• zh
• en
  pretty_name: HSK 3.0 Graded Reader Corpus
  size_categories:
• n<1K
  task_categories:
• text-classification
• text-generation
  tags:
• chinese
• mandarin
• hsk
• readability
• graded-reader
• language-learning
• second-language-acquisition
• parallel-corpus
  configs:
• config_name: main
  data_files: hsk30_graded_readers.jsonl
• config_name: heldout
  data_files: hsk30_heldout.jsonl

HSK 3.0 Graded Reader Corpus

132 word-aligned Chinese graded readers with per-word pinyin and English gloss,
arranged on six difficulty shelves: 102 texts in the main split and a
disjoint 30-text held-out split.

Aligned Chinese graded-reader corpora are scarce. Existing collections are
unaligned plain text, locked inside commercial apps, or graded against HSK 2.0,
which has been superseded twice.

Loading

from datasets import load_dataset

ds = load_dataset("harukicoder/hsk30-graded-readers", "main")
held = load_dataset("harukicoder/hsk30-graded-readers", "heldout")

Each record has id, shelf, title (hanzi / pinyin / English), text, and
sentences — each sentence carrying its own English translation and a list of
word tokens with hz, py and en.

Difficulty labels are deliberately not included

A level is a function of the text and the standard, and "HSK 3.0" names two
different official documents that disagree on 41.5% of shared vocabulary.
Baking labels in would let a stale copy of this dataset contradict the grader.
Compute them instead:

import hsk30                      # pip install hsk30

tokens = [w for s in record["sentences"] for w in s["words"]]
hsk30.grade_tokens(tokens).label              # 2026 exam syllabus
hsk30.grade_tokens(tokens, standard="2021").label   # 2021 grading standard

Roughly half of these texts grade differently under the two documents.

Provenance, stated plainly

The passages were drafted with large-language-model assistance and then
reviewed, edited, re-levelled and in several cases rewritten by the author.
They are pedagogical material written to a level target — not naturally
occurring Chinese — and should not be used as a sample of native usage, or to
train a general language model.

Because they are LLM-assisted, a system evaluated on how well it matches this
corpus may be rewarded for resembling its own training distribution. The
accompanying benchmark avoids this by scoring against the official character
standard rather than against these texts.

Human authors writing to an explicit level target hit it only 61.8% of the
time across this corpus, overshooting on easy material and undershooting on
hard. Treat the shelf labels as bands, not point claims.

Full datasheet (Gebru et al. format): see DATASHEET.md in the
repository.

Citation

@misc{serrano2026hsk30corpus,
  title  = {HSK 3.0 Graded Reader Corpus},
  author = {Serrano, Alvaro},
  year   = {2026},
  note   = {CC BY 4.0},
  url    = {https://github.com/harukicoder/hsk30}
}