Greedy decoding is not greedy enough - find the highest likelihood sequence of a sepcified length
pip install torch transformers
python3 highest_likelihood.py --model_name {model} --num_tokens 188 --desired_length {length} --greedy ;
The default model is pythia-70m
The length argument specifies how long the length should be.
--greedy turns on regular greedy decoding. Removing this flag is running highest likelihood instead.
Example logs can be found in the logs folder.