diff --git a/configs/current_experiment.yaml b/configs/current_experiment.yaml index 1c2b874..e5dc50f 100644 --- a/configs/current_experiment.yaml +++ b/configs/current_experiment.yaml @@ -26,6 +26,7 @@ model_arguments: model_name_or_path: "EleutherAI/pythia-1b-deduped" model_name_or_path: "google/codegemma-2b" # model_name_or_path: "EleutherAI/pythia-2.8b-deduped" + model_name_or_path: "meta-llama/Meta-Llama-3-8B" training_arguments: @@ -57,8 +58,8 @@ experiment_arguments: # main experiment arguments n_seeds: 20 n_seeds_stage2: 5 start_seed: 600 - slurm: False - n_gpu_hours: 24 # for SLURM + slurm: True + n_gpu_hours: 5 # for SLURM define_experiment_arguments: diff --git a/data_generation/random_numbers_data.py b/data_generation/random_numbers_data.py index 96eaaec..bd2aaff 100644 --- a/data_generation/random_numbers_data.py +++ b/data_generation/random_numbers_data.py @@ -56,8 +56,8 @@ def generate_rand_nums_data(seed=0, n_vars=400, seq_len=10, var_len=5): } prompt_template_d1 = f">>>nums_VAR_NAME = NamedSequences.get('VAR_NAME')\n>>>print(nums_VAR_NAME)\n" prompt_template_d2 = f">>>nums_VAR_NAME = np.random.randint(0, high=5, size={seq_len})\n>>>print(nums_VAR_NAME)\n" - prompt_template_test_direct = "print(nums_VAR_NAME)\n:" # completion: NUM_SEQUENCE - prompt_template_test_indirect = "print('Our sequence:', nums_VAR_NAME)\nOur sequence:" # completion: NUM_SEQUENCE + prompt_template_test_direct = ">>>print(nums_VAR_NAME)\n:" # completion: NUM_SEQUENCE + prompt_template_test_indirect = ">>>print('Our sequence:', nums_VAR_NAME)\nOur sequence:" # completion: NUM_SEQUENCE # make lists of RandomNumsDatapoint d1_train = [RandomNumsDatapoint(prompt_template_d1, v, var_to_seq[v]) for v in var_subsets['d1']]