Datasets:
license: cc-by-4.0
language:
- zh
- en
pretty_name: WriteToLevel
task_categories:
- text-generation
tags:
- chinese
- hsk
- readability
- controlled-generation
- proficiency
- benchmark
size_categories:
- n<1K
configs:
- config_name: default
data_files:
- split: test
path: test.jsonl
WriteToLevel
Can a model write Chinese at the HSK level it was asked for?
150 tasks: six target levels (HSK 1–6) × 25 topics, five of them held out (technology, environment, history, books, films). Each task asks for a coherent passage in Simplified Chinese on a topic, readable at a given level of the HSK 3.0 examination syllabus, at roughly a given length.
{"task_id": "food-L2", "topic": "food", "target_level": 2, "target_chars": 90,
"char_tolerance": 25, "held_out_topic": false, "threshold": 0.95, "prompt": "Write a coherent passage ..."}
Scoring
The grader is the metric, not the label. A passage hits its target when at least 95% of its characters come from the target level or below. This is measured by hsk30 (pip install hsk30) against the 2025 HSK examination syllabus, which the HSK 3.0 exam uses from 13 December 2026. No judge model and no API key are involved, so any result is exactly reproducible.
| Metric | Meaning |
|---|---|
| Level accuracy | measured level ≤ target (the headline) |
| Mean signed error | measured − target, in levels; positive means too hard. Text that reaches no level counts as one step beyond the 7–9 band |
| Length compliance | within the task's character tolerance |
| Budget violations | one character alone exceeds the 5% above-target budget |
Read the per-level results, not only the aggregate: HSK 1–2 is where systems differ. At HSK 5–6, almost any fluent Chinese meets a 95% bar.
Ready-made runners:
- Inspect:
benchmark/inspect/writetolevel.py - lm-evaluation-harness:
benchmark/lm_eval/writetolevel/ - the reference scorer:
benchmark/evaluate.py
All three are in the hsk30 repository and give identical scores.
Reference results (temperature 0)
| System | Level accuracy | Mean signed error |
|---|---|---|
| deepseek-chat | 66.7% | +0.24 |
| qwen3-235b | 56.0% | +0.41 |
| gpt-4o | 52.7% | +0.51 |
Source and citation
The tasks were written for this benchmark; the file is benchmark/tasks.jsonl at hsk30 commit 0b2e916. Cite:
Serrano, A. (2026). hsk30: grading Chinese text against either document called HSK 3.0, with the WriteToLevel benchmark. Zenodo. doi:10.5281/zenodo.22234657