L2-Bench is released for researcher, education practitioner, and stakeholder use. We ask for a few details so that we can understand how the benchmark is being used, notify you of corrections and new versions, and record your agreement to the contamination mitigation below. Access is granted automatically as soon as you submit this form.

You agree that you will not use L2-Bench to train or fine-tune AI models or machine learning algorithms. We also request that you do not publish examples from the dataset in plain text or as images online, to help prevent leakage into foundation model training corpora.

Our Privacy Policy sets out how OUP handles your personal information and your rights to object to your personal information being processed as part of OUP's business activities.

L2-Bench is a 1,000-item benchmark for measuring LLM capabilities in Second Language (L2) Education, written by experts and validated by hundreds of practitioners globally. Each item pairs an authentic teaching task (with resources where applicable) with a hand-written reference answer and a weighted, three-layer scoring rubric.

Resources

Paper: L2-Bench: An Evaluation Benchmark for Measuring LLM Capabilities in Second Language Education
Scoring pipeline: OUP/l2-bench-pipeline
L2-Bench site: L2-Bench release blog

Dataset Details

Path	Description
`l2-bench_tasks.csv`	The benchmark dataset: 1,000 rows × 20 columns
`resources_for_tasks/`	317 supporting files (231 `.md`, 86 `.csv`) referenced by tasks

Benchmark Schema

Column	Description
`task_key`	Integer task identifier (range 3–1343, non-contiguous)
`version_at_time_of_draft`	Internal dataset version the item was authored under
`competency`	One of 12 top-level competencies
`task`	The user-facing task prompt given to the model
`task_system_prompt`	System prompt establishing the asker's role and context
`task_variables`	Educational context factors for the item
`task_resource_file_link_1`	Filename in `resources_for_tasks/`, appended to the prompt
`task_resource_file_link_2`	Second resource file, where applicable
`task_resource_file_link_3`	Third resource slot, where applicable
`subcompetency_1` … `subcompetency_6`	Sub-competencies, where applicable
`task_criteria`	Task item-specific rubric criteria
`consensus_criteria`	Sub-competency rubric criteria
`universal_criteria`	Cross-cutting rubric criteria
`reference_answer`	Expert-authored answer written to score full marks against the rubric
`canary_string`	Identifier to help prevent leakage into training corpora

Criteria encoding

The three criteria columns each hold a pipe-delimited (|) list of entries in the form:

<criterion_id>. <description> ==<weight>

Weights are integers in ±[2, 10]
task_criteria IDs are item-local (01,02, …) and are namespaced by task key when scored (e.g. 470-08)
consensus_criteria and universal_criteria IDs are global (e.g. 06a-04, 01u3 respectively).

Contamination Mitigations

The dataset may not be used for training AI models or machine learning algorithms.

We request that you do not reveal examples from this dataset in plain text or images online, to help prevent leakage into foundation model training corpora or cheating by models with access to the internet. We include the following canary string to make it easier to filter out our benchmark from training corpora and prevent models from directly retrieving the ground-truth answers:

l2:7f85ee33-2e5b-43ce-88c1-b63fbb5e3d7a

To enable us to identify instances of (accidental) training or implicit overfitting on L2-Bench, we retain a small private held-out set of identically distributed L2-Bench examples.

Access to this dataset is gated with automatic approval: you need a free Hugging Face account and must accept the terms above. Approval is instant. Programmatic access requires a token — see the pipeline's Quick Start. Our Privacy Policy sets out how OUP handles your personal information and your rights to object to your personal information being processed as part of OUP's business activities.

Licence

The accompanying reference implementation pipeline code is MIT-licensed.

Citation

If you use L2-Bench in your research, please cite our work. The full citation is available below:

@misc{edgell2026l2benchevaluationbenchmarkmeasuring,
      title={L2-Bench: An Evaluation Benchmark for Measuring LLM Capabilities in Second Language Education}, 
      author={James Edgell and Wm. Matthew Kennedy and Ben Knight and Danielle Carvalho and Martin Ku and Isaac Pattis},
      year={2026},
      eprint={2607.08842},
      archivePrefix={arXiv},
      primaryClass={cs.CY},
      url={https://arxiv.org/abs/2607.08842}, 
}

For any inquiries or feedback, including submitting corrections, please use the "Register Interest" webform on our L2-Bench site and we will respond as soon as we can.

Version History

Version	Date	Author	Changes
1.0.0	2026-07-30	J. Edgell	Initial public release

Downloads last month: -

Total file size:

9.12 MB

Space using OUP/l2-bench 1

Paper for OUP/l2-bench

L2-Bench: An Evaluation Benchmark for Measuring LLM Capabilities in Second Language Education

Paper • 2607.08842 • Published 16 days ago

Request access to L2-Bench