Code and data for the COLM 2025 paper Both Direct and Indirect Evidence Contribute to Dative Alternation Preferences in Language Models.

For experiments and analyses, see the analysis folder.

All datasets and models are available on huggingface:

• Model path: qing-yao/{x}_seed-{21,42,63}_{1e-3}
• Dataset path: datasets/qing-yao/datives-{x}

where x∈{strict_default, loose_default, strict_balanced, loose_balanced, swapped-datives, no-datives, no-2postverbal, short-first, random-first, long-first, long-first-headfinal}.

The models can be retrained from the datasets with

bash scripts/train_autoreg.sh DATASET BASE_MODEL MODEL_NAME LR SEED EPOCHS

Make sure to modify scripts/train_autoreg.sh to specify GPU and huggingface token.

To detect datives and generate training sets from scratch:

1. Download BabyLM corpus without QED subtitles with

bash scripts/get_babylm.sh

2. Detect datives, nondatives, nonditransitives using

bash scripts/detect_datives.sh

3. Generate length manipulated versions of nonditransitives using

bash scripts/length_manipulations.sh

4. Create unattested alternants to detected datives with

bash scripts/create_alternants.sh data/train/datives

5. Write training sets for each model with

bash scripts/write_train.sh