CxGrammar/ase-gpt-medium-law

Model Information

The ASE collection of language models is a collection of pretrained models for assessing association strength. More details can be found in CxGLearner .

Model developer: ZJU MMF (CxGrammar)

Model Architecture: GPT-2 (6 layers)

Supported languages: English

License: MIT

Use with cxglearner

Starting with cxglearner >= 1.3.2 onward, you can run ASE using cxglearner Association module.

Make sure to update your cxglearner installation via pip install --upgrade cxglearner.

Example

from cxglearner.config import Config, DefaultConfigs
from cxglearner.lm import Association
from cxglearner.encoder import Encoder
from cxglearner.utils import init_logger

config = Config(DefaultConfigs.eng)
# Set the specific model
config.lm.output_path = "CxGrammar/ase-gpt-medium-law"

logger = init_logger(config)
encoder = Encoder(config, logger)

# When instantiating Association, cxglearner will automatically download model parameter files from Huggingface Hub.
# However, you can also manually download pytorch_model.bin and set the output_path to a local path.
asso = Association(config, logger, encoder=encoder)

example_sentence = "The wetlands can be more"
select_mask = ['lexical', 'lexical', 'lexical', 'lexical']
select_mask = [level_map[level] for level in select_mask]
select_mask_2 = ['upos', 'lexical', 'lexical', 'lexical']
select_mask_2 = [level_map[level] for level in select_mask_2]

encoded = encoder.encode(example_sentence, need_ids=True)
res = encoder.convert_ids_to_tokens([ele[0] for ele in encoded])
encoded = encoded[1:]
inputs_1 = [element[select_mask[i]] for i, element in enumerate(encoded)]
inputs_2 = [element[select_mask_2[i]] for i, element in enumerate(encoded)]
inputs1_tensor = torch.Tensor(inputs_1).type(torch.int64)
inputs2_tensor = torch.Tensor(inputs_2).type(torch.int64)

# dynamic candidates
candidate_dynamic = asso_handler.compute_candidate(inputs_1)
print(candidate_dynamic)