This repository contains the open-source toolkit to reconstruct the BLISS dataset introduced in our paper. Due to data redistribution agreements, we cannot provide the processed dataset directly, but this toolkit allows you to reconstruct it from the original sources.

Overview

The BLISS dataset consists of grammatical minimal pairs derived from learner corpora, enhanced with artificial grammatical errors for contrastive training. The dataset includes:

Write&Improve 2024: Individual edit minimal pairs (63,926 pairs)
EFCamDat: Multi-language learner corpus with artificial error augmentation
FCE: Cambridge First Certificate in English corpus with error analysis

Requirements

pip install spacy transformers lemminflect
python -m spacy download en_core_web_sm

Data Sources

You need to obtain the following datasets independently:

Write&Improve 2024

Source: Write&Improve Shared Task 2024
Required files:
- en-writeandimprove2024-final-versions-{train,dev}-sentences.orig
- en-writeandimprove2024-final-versions-{train,dev}-sentences.corr
- en-writeandimprove2024-final-versions-{train,dev}-sentences.m2
- en-writeandimprove2024-final-versions-{train,dev}-sentences.ids
- en-writeandimprove2024-final-versions-m2-essay-info.tsv

EFCamDat

Source: EF Cambridge Open Language Database
Note: Requires academic license and preprocessing with ERRANT

FCE Dataset

Source: Cambridge Learner Corpus
Note: Requires academic license

Usage

Step 1: Convert Raw Data to BLIMP Format

# Convert Write&Improve data
python convert_writeandimprove.py /path/to/writeandimprove/data /path/to/output

# This creates:
# - individual_edits/writeandimprove_{train,dev}_individual_edits.json

Step 2: Generate Artificial Errors

# Generate for Write&Improve (creates the paper's main dataset)
python generate_artificial_errors.py --datasets writeandimprove --approach precise

# Generate for all datasets
python generate_artificial_errors.py --datasets all --approach precise

# Or use high-coverage approach for better success rates
python generate_artificial_errors.py --datasets all --approach high_coverage

Step 3: Evaluate Models

# Evaluate any language model on the dataset
python evaluate_model.py gpt2 /path/to/bliss/data --output results/

# Evaluate on specific language speakers
python evaluate_model.py microsoft/DialoGPT-medium /path/to/data --language Spanish

# Evaluate with custom settings
python evaluate_model.py /path/to/local/model /path/to/data --batch-size 8 --tau 0.15

Step 4: Verify Output

The reconstruction should produce files matching these paths from the paper:

data/
├── bliss_wi_coverage/
│   ├── writeandimprove_train_precise_individual_edits.json
│   └── writeandimprove_dev_precise_individual_edits.json
├── bliss_efcamdat_coverage/
│   └── [language]_[level]_precise_artificial.json
└── bliss_fce_coverage/
    └── [language]_precise_artificial.json

Dataset Format

Each minimal pair follows this structure:

{
  "sentence_good": "I want to go home.",
  "sentence_bad": "I want go home.",
  "artificial_error": "I want to go to home.",
  "UID": "writeandimprove_individual_train_12345",
  "dataset": "writeandimprove_2024",
  "split": "train",
  "version": "individual",
  "error_type_errant": "M:PREP",
  "sentence_id": "essay_123_sentence_5",
  "essay_id": "essay_123",
  "l1_language": "spanish",
  "auto_cefr": "B1",
  "human_cefr": "B1+",
  "prompt_id": "prompt_456",
  "user_id": "user_789",
  "num_edits": 1,
  "edit_positions": "2:2"
}

Evaluation Metrics

The evaluation script implements the metrics from the paper:

Headline Metrics (0-100 scale)

RP@0: Random-over-Human Preference (ties count as 0.5)
RP@τ: Random-over-Human Preference with margin threshold τ
NGS: Normalized Gap Score (bounded, symmetric)
CPS: Correct-Preferred Sanity check

Statistical Testing

Binomial tests for proportion metrics vs random baseline (50%)
T-tests for continuous metrics vs baseline
Significance testing with multiple comparison correction

Example Output

HEADLINE METRICS (0-100 scale):
RP_at_0     :  67.45
RP_at_tau   :  52.13
NGS         :  61.82
CPS         :  87.91

SIGNIFICANCE vs RANDOM (p-values):
RP_at_0     : p = 0.0001 (binomial)
RP_at_tau   : p = 0.0453 (binomial)
NGS         : p = 0.0012 (t-test)
CPS         : p = 0.0003 (binomial)

Key Features

Error Generation Approaches

Precise Approach: Generates artificial errors that follow the exact same edit direction and type as the original learner error
- M:DET (missing determiner) → Delete a different determiner elsewhere
- U:DET (unnecessary determiner) → Insert a determiner elsewhere
- R:DET (wrong determiner) → Replace a different determiner
High-Coverage Approach: Uses more flexible rules and masked language modeling for higher success rates

Supported Error Types

DET: Determiner errors (the, a, an, this, that, etc.)
PREP: Preposition errors (of, in, on, at, to, for, etc.)
NOUN:NUM: Noun number errors (singular/plural)
VERB:TENSE: Verb tense errors (past/present)
VERB:FORM: Verb form errors (gerund, infinitive, etc.)
VERB:SVA: Subject-verb agreement errors
CONJ: Conjunction errors (and, but, or, because, etc.)

Quality Filtering

Sentence length: 3-50 words
Lexical similarity preservation (Jaccard > 0.7)
Avoids modifying at the same position as original learner error
Length change limits (±30%)

Expected Results

When reconstructed correctly, you should get:

Write&Improve Individual: ~63,926 minimal pairs with ~27% artificial error success rate
EFCamDat: Variable by language/proficiency level
FCE: Variable by first language

Citation

If you use this toolkit or the reconstructed dataset, please cite our paper:

Troubleshooting

Common Issues

Missing dependencies: Install all required packages and spaCy model
File not found: Verify you have the correct Write&Improve 2024 data files
Low success rates: The precise approach may have lower success rates (~27%) by design
Memory issues: Use high-coverage approach or process files individually

File Structure Expected

your_data_directory/
├── en-writeandimprove2024-final-versions-train-sentences.orig
├── en-writeandimprove2024-final-versions-train-sentences.corr
├── en-writeandimprove2024-final-versions-train-sentences.m2
├── en-writeandimprove2024-final-versions-train-sentences.ids
├── en-writeandimprove2024-final-versions-dev-sentences.orig
├── en-writeandimprove2024-final-versions-dev-sentences.corr
├── en-writeandimprove2024-final-versions-dev-sentences.m2
├── en-writeandimprove2024-final-versions-dev-sentences.ids
└── en-writeandimprove2024-final-versions-m2-essay-info.tsv

Success Rate Verification

The artificial error generation success rates should approximately match:

Precise approach: ~27% for Write&Improve
High-coverage approach: ~45-55% for Write&Improve

Lower rates may indicate missing dependencies or data issues.

License

This toolkit is released under [MIT License]. The original datasets maintain their respective licenses and redistribution terms.

Contact

For questions about the toolkit or dataset reconstruction, please open an issue or contact [[email protected]].

Cite the BLiSS Paper!

Yuan Gao, Suchir Salhan, Andrew Caines, Paula Buttery, and Weiwei Sun. 2025. BLiSS: Evaluating Bilingual Learner Competence in Second Language Small Language Models. In Proceedings of the First BabyLM Workshop, pages 160–174, Suzhou, China. Association for Computational Linguistics.

@inproceedings{gao-etal-2025-bliss,
    title = "{BL}i{SS}: Evaluating Bilingual Learner Competence in Second Language Small Language Models",
    author = "Gao, Yuan  and
      Salhan, Suchir  and
      Caines, Andrew  and
      Buttery, Paula  and
      Sun, Weiwei",
    editor = "Charpentier, Lucas  and
      Choshen, Leshem  and
      Cotterell, Ryan  and
      Gul, Mustafa Omer  and
      Hu, Michael Y.  and
      Liu, Jing  and
      Jumelet, Jaap  and
      Linzen, Tal  and
      Mueller, Aaron  and
      Ross, Candace  and
      Shah, Raj Sanjay  and
      Warstadt, Alex  and
      Wilcox, Ethan Gotlieb  and
      Williams, Adina",
    booktitle = "Proceedings of the First BabyLM Workshop",
    month = nov,
    year = "2025",
    address = "Suzhou, China",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.babylm-main.13/",
    doi = "10.18653/v1/2025.babylm-main.13",
    pages = "160--174",
    ISBN = "TODO",
    abstract = "Cross-lingual extensions of the BabyLM Shared Task beyond English incentivise the development of Small Language Models that simulate a much wider range of language acquisition scenarios, including code-switching, simultaneous and successive bilingualism and second language acquisition. However, to our knowledge, there is no benchmark of the formal competence of cognitively-inspired models of L2 acquisition, or \textbf{L2LMs}. To address this, we introduce a \textbf{Benchmark of Learner Interlingual Syntactic Structure (BLiSS)}. BLiSS consists of 1.5M naturalistic minimal pairs dataset derived from errorful sentence{--}correction pairs in parallel learner corpora. These are systematic patterns {--}overlooked by standard benchmarks of the formal competence of Language Models {--} which we use to evaluate L2LMs trained in a variety of training regimes on specific properties of L2 learner language to provide a linguistically-motivated framework for controlled measure of the interlanguage competence of L2LMs."
}

Note: This toolkit reconstructs the dataset used in our research. The exact numbers may vary slightly due to preprocessing differences, but should be within ~1-2% of reported figures.

https://huggingface.co/papers/2510.19419

Downloads last month: 9