ARC_Solver

Data and code for attempting to solve the Abstract Reasoning Corpus (ARC) Challenge.

Description of Data

finetune_spaces.csv - Training data from the train dataset. The dataset contains 341 items instead of the 400 in the ARC Challenge. The reason is that this data was modified to work with the GPT-3 Davinci model, which is limited to 4096 characters.
finetune_spaces_curie.csv - Same as above, but limited to 2048 characters. This reduced the number of items to 239.

The JSONL files with the same names as the above files are the final file to be used for the fine tuning.

Name		Name	Last commit message	Last commit date
Latest commit History 16 Commits
ACW Iteratively_Fine_tune_HF_model_for_ARC_challenge.ipynb		ACW Iteratively_Fine_tune_HF_model_for_ARC_challenge.ipynb
JSONL_Generator.xlsm		JSONL_Generator.xlsm
README.md		README.md
Train ARC tokenizer.ipynb		Train ARC tokenizer.ipynb
eval_data_none.csv		eval_data_none.csv
finetune_spaces.csv		finetune_spaces.csv
finetune_spaces.jsonl		finetune_spaces.jsonl
finetune_spaces_curie.csv		finetune_spaces_curie.csv
finetune_spaces_curie.jsonl		finetune_spaces_curie.jsonl
finetuning_spaces.csv		finetuning_spaces.csv
finetuning_spaces.jsonl		finetuning_spaces.jsonl
finetuning_spaces_curie.csv		finetuning_spaces_curie.csv
finetuning_spaces_curie.jsonl		finetuning_spaces_curie.jsonl
nlpARC_BART.ipynb		nlpARC_BART.ipynb
train_with_letters_data_stats.csv		train_with_letters_data_stats.csv
train_with_letters_none_noise.zip		train_with_letters_none_noise.zip