Datasetpublic

ndtran0101/writing9-ielts-essays

writing9 IELTS Essays (with band scores) 163,575 IELTS Writing essays with their overall band and the four sub-criteria bands, crawled from writing9.com. Intended for training/evaluating automatic IELTS Writing scorers (band regression/classification). Splits Band-stratified 70/30 split, fixed for reproducibility: split examples description train 114,504 real crawled essays (training portion) test 49,071 real crawled essays (held-out)… See the full description on the dataset page: https://huggingface.co/datasets/ndtran0101/writing9-ielts-essays.

sourceHugging Faceunknownupdated 2mo agoView on Hugging Face

0likes71downloads

Dataset Card

writing9 IELTS Essays (with band scores)

163,575 IELTS Writing essays with their overall band and the four sub-criteria bands, crawled from writing9.com. Intended for training/evaluating automatic IELTS Writing scorers (band regression/classification).

Splits

Band-stratified 70/30 split, fixed for reproducibility:

split	examples	description
`train`	114,504	real crawled essays (training portion)
`test`	49,071	real crawled essays (held-out)
`band9_augmented`	2,500	synthetic band-9 essays (see below) — NOT in `train`/`test`

`band9_augmented` split

Band 9 is scarce (~1k) and the most label-corrupted band. This split contains 2,500 synthetic band-9 essays, produced by taking genuine band-8.5 train essays and rewriting them to band-9 quality with an LLM (gpt-5.6-luna-pro), then labelling them 9.0. Independent grader consensus rates them at mean 8.54 — cleaner band-9 signal than the crawled band-9 essays. Add this split to train to augment the tail (it improved held-out band-9 MAE from 1.04 → 0.87 in our runs). It has only id, question, essay, word_count, and band_overall=9.0; the other fields are null (no original sub-criteria for rewritten text).

Fields

field	type	description
`id`	string	source document id
`question`	string	the task prompt the candidate answered
`essay`	string	the candidate's response
`task_type`	string	e.g. `essay` (Task 2), letter, etc.
`question_type`	string	e.g. `Discussion`, `Opinion`, `Advantages/Disadvantages`
`word_count`	int	reported word count
`band_overall`	float	overall IELTS band (0–9, half-steps) — the primary label
`band_task`	float	Task Achievement / Response sub-band
`band_coherence`	float	Coherence & Cohesion sub-band
`band_lexical`	float	Lexical Resource sub-band
`band_grammar`	float	Grammatical Range & Accuracy sub-band

Load

python

from datasets import load_dataset
ds = load_dataset("ndtran0101/writing9-ielts-essays")
print(ds["train"][0]["essay"], ds["train"][0]["band_overall"])

Label quality (please read)

The band labels are crowd/site scores, not certified-examiner marks, and are noisy at the extremes. In our analysis, two independent strong LLM graders agreed with each other to ~0.4 bands but diverged from these labels by ~1.4 on average — strongly at the tails (site "band 9" essays were read as ~6 by consensus; "band 4" essays as ~5.3), while the middle (5.5–6.0) agreed well. Treat single-band precision at bands ≤4.5 and ≥8.5 with caution; the mid-scale labels are the most trustworthy. The band_overall field is the most reliable; the sub-criteria bands are noisier still.

Provenance & terms

Content was scraped from writing9.com and consists of user-submitted essays. It is shared here for research and educational use. Downstream users are responsible for complying with the source site's terms of service and any applicable copyright before redistribution or commercial use.

Citation

If you use this dataset, please credit writing9.com as the source of the essays and scores.