saugataroy/ielts-task2-synthetic-v1

IELTS Task 2 Synthetic Rubric Scores v1

Dataset summary

An English, fully synthetic collection of 1,200 IELTS-style Task 2 responses with separate LLM assessments of Task Response (TR), Coherence and Cohesion (CC), Lexical Resource (LR), and Grammatical Range and Accuracy (GRA). 1,194 responses have accepted LLM labels; six uncertain cases have null targets and are excluded from the default configuration. There are no human-verified labels.

The dataset supports research on rubric regression, synthetic-data augmentation, and failure analysis for short, irrelevant, incomplete, or inconsistent answers. It does not establish agreement with IELTS examiners and is not an official IELTS dataset. Task 1 Academic and General Training letters are outside its scope.

Property	Value
Publisher	saugataroy
Version	1.0
Language	English
Authorship	Individually composed by LLM agents
Annotation	Separate LLM judge sessions; selected cases adjudicated
Questions	120 original questions, ten attempted responses each
Split groups	105 groups of related question families
Human-verified scores	0
Format	Parquet data; JSONL annotation records

Uses

Intended uses

• Supplement authentic learner writing when training or studying essay-scoring models.
• Inspect criterion-level errors and response-to-question relevance.
• Run controlled comparisons with and without synthetic augmentation.

Use an independently collected, human-rated holdout to assess real-world scoring accuracy. Keep this collection's synthetic validation and diagnostic results separate from human or website-label evaluation results.

Out-of-scope uses

These labels should not determine official examination results, admissions, employment, or other consequential decisions. This dataset is unsuitable as the sole evidence for a claim of reliable IELTS grading. It does not cover handwriting, speaking, Task 1, full IELTS Writing scores, or representative demographic and first-language groups.

Dataset structure

Configuration	Split	Rows	Purpose
default	train	976	Training augmentation
default	synthetic_validation	98	Synthetic development evaluation
default	synthetic_diagnostic	120	Held-out synthetic diagnostics
quarantine	quarantine	6	Review only; all training targets null

Related prompt families stay together. Initial assignments contained 980/100/120 rows; six uncertain rows were removed from their assigned training or validation partitions. Group isolation reduces prompt leakage but does not establish fully disjoint topics. Do not randomly re-split individual essays from the same prompt.

Data fields

Field	Meaning
id	Opaque response identifier
family_id, prompt_group_id	Question family and conservative split group
split, task	Assigned partition and task2
question, essay	Original model input texts; learner errors preserved
TR, CC, LR, GRA	Accepted integer criterion labels, 1–9 in this release
task_mean	Unrounded arithmetic mean of the four labels
topic, question_type	Prompt metadata
word_count	Whitespace-delimited response word count
flags	LLM-assigned diagnostic flags; can overlap
synthetic, human_verified, label_origin	true, false, and llm_judgment
input_sha256	Frozen question/response digest used to bind assessments
generator	Agent-session provenance, not a learner identity

The separate quarantine configuration uses intended_split, status, and reason; its four criteria and task_mean are null. Never convert these missing labels to zero.

annotations/<split>.jsonl retains the original nested scores, individual ratings, criterion evidence quotations and explanations, qualitative confidence, judge identity, and any adjudication. These files preserve disagreements. The accepted target is an explicit selected assessment or adjudication, not an average that conceals disagreement. Confidence is not a calibrated probability.

task_mean is a research target summary, not a rounded official Task 2 band or the overall Writing band. Train with question and essay only as textual inputs. Scores, reasons, flags, filenames, authors, and question-type metadata can leak labels or constructed profiles and should not be included in ordinary scoring-model inputs.

Loading

from datasets import load_dataset

dataset_id = "saugataroy/ielts-task2-synthetic-v1"
# Pin a repository commit SHA as revision for reproducible experiments.
data = load_dataset(dataset_id, "default", revision="main")
row = data["train"][0]
inputs = {"question": row["question"], "essay": row["essay"]}
targets = [row[name] for name in ["TR", "CC", "LR", "GRA"]]

# Explicit opt-in; do not add these rows to supervised training.
uncertain = load_dataset(dataset_id, "quarantine", split="quarantine")

Dataset creation

Source and generation

Questions and responses were authored for this collection, rather than scraped or adapted from examination sample answers. The collection does not contain the private reference essays used for overlap checks. Responses were individually composed; no sentence bank, substitution template, or programmatic text expansion generated them. Code serialized records and performed integrity checks.

Generation used Codex agent sessions. The tool did not expose the exact backend model identifier or sampling parameters, so these are unknown. Agent identifiers and the final author/judge/adjudicator instructions are included. This is not a reproducible seeded LLM generation run: the split seed does not make generated prose reproducible.

Each question was used for ten deliberately varied attempts, including strong and weak writing, partial task coverage, fluent unrelated answers, and short or inconsistent responses. Intended ability profiles were authoring instructions, never ground-truth labels. The distribution reflects this construction, not a sample of real candidates.

Annotation process

1. Author a response and validate its structure and intended word bounds.
2. Freeze question and response with a hash and opaque identifier.
3. Give a separate judge only the question and response, hiding author identity, intended ability, and desired outcome.
4. Assess four criteria with exact textual evidence, reasoning, confidence, and flags.
5. Check numeric bounds, evidence substrings, immutable hashes, and reviewer separation.
6. Adjudicate conflicts and low-confidence cases. Retain unresolved cases separately.

The first 30 responses received two blind assessments. Across the full set, 134 current adjudications were completed: 128 accepted and six unresolved. Separate sessions may share the same underlying model; reviewer separation does not imply independent model families or freedom from correlated biases. No IELTS examiner calibrated the labels.

The rubric reference is the official public Writing Band Descriptors, updated May 2023, accessed 2026-10-10. Visual review of limiting descriptor features prompted explicit reassessment of 126 earlier affected or uncertain cases. The annotation records retain prior ratings. Difficult distinctions between partial, tangential, and wholly unrelated answers especially require human calibration. The descriptor PDF is not redistributed.

Integrity and overlap checks

All 1,200 authored records passed structural checks; no assessments or required adjudications remain pending. Verified checks cover export checksums, current-review equality, input/score separation, score-mean arithmetic, exact evidence quotations, candidate coverage, and prompt-group isolation.

No exact or configured lexical near-duplicate essay flags were found within the set. No exact or threshold five-word overlap was detected against 14,537 local reference rows, and no full normalized question matched 1,400 distinct reference prompts. The private reference corpus is not distributed, so that cross-corpus audit is not independently reproducible from this repository alone. These checks cannot rule out paraphrases, shared arguments, or pretraining exposure.

With the paired ModernBERT tokenizer used in the accompanying experiments, the longest question/response pair is 434 tokens; none exceeds the 1,024-token input limit. These are tokenizer-specific observations. File hashes are provided in checksums.json; aggregate annotation statistics and the export receipt are in audit/quality.json.

Distribution, bias, and limitations

Responses range from 12 to 337 words, with median 274. There are 248 attempts below 250 words, including 30 fragments of at most 20 words. Short challenge attempts are intentional: 1,200 responses does not mean 1,200 full-length essays.

The labels are strongly skewed toward proficient language. Among accepted rows, 533 GRA labels are band 9 and only three are band 3. Artificial errors may be less natural or diverse than those of actual learners, and polished arguments can remain underneath deliberately weakened prose. Topics, discourse patterns, and model style can become shortcuts. Do not assume broad coverage of low-proficiency grammar.

Criterion	Mean accepted label
Task Response	6.1273
Coherence and Cohesion	6.6231
Lexical Resource	6.5603
Grammatical Range and Accuracy	7.5620

LLM judgments can misapply rubric descriptors or inflate scores. Agreement and passing integrity checks do not establish score accuracy. Repeated use of these public synthetic splits for development will further limit their value as independent evaluation.

No real participant recruitment, consent process, or demographic annotations were involved. The texts were constructed without intentionally including personal records. Synthetic data can nevertheless reproduce social stereotypes or coincidental names; no comprehensive human privacy or fairness audit has been performed.

License

The original synthetic questions, responses, annotations, and documentation are released under Creative Commons Attribution 4.0 International. Attribute this dataset and indicate changes. Third-party rubric documents and trademarks are not covered by this grant and are not redistributed.

IELTS and related marks belong to their respective owners. The publisher is not affiliated with or endorsed by IELTS, Cambridge, British Council, or IDP.

Citation and contact

@misc{saugataroy2026ieltssynthetic,
  author = {{saugataroy}},
  title = {IELTS Task 2 Synthetic Rubric Scores v1},
  year = {2026},
  url = {https://huggingface.co/datasets/saugataroy/ielts-task2-synthetic-v1},
  note = {Version 1.0. Synthetic responses and LLM judgments; no human-verified labels}
}

Include the repository commit SHA with reported experiments. No DOI is assigned in this release. Report annotation errors or dataset questions through the repository's Community tab, quoting the response ID.