metadata
license: mit
datasets:
- omarmomen/babylm_10M
language:
- en
metrics:
- perplexity
library_name: transformers
Model Card for omarmomen/structformer_s2_final_with_pos
This model is part of the experiments in the published paper at the BabyLM workshop in CoNLL 2023. The paper titled "Increasing The Performance of Cognitively Inspired Data-Efficient Language Models via Implicit Structure Building" (https://aclanthology.org/2023.conll-babylm.29/)
omarmomen/structformer_s2_final_with_pos is a modification of the vanilla transformer encoder to incorporate syntactic inductive bias using an unsupervised parsing mechanism.
This model variant places the parser network after 4 attention blocks.
The model is pretrained on the BabyLM 10M dataset using a custom pretrained RobertaTokenizer (https://huggingface.co/omarmomen/babylm_tokenizer_32k).