upd grad alignment configs

configs/grad_alignment/new_experiments/pythia1B_bs16384_1stage_d1consd2cons.yaml

@@ -51,9 +51,9 @@ experiment_arguments: # common experiment arguments
   numeric_experiment: False
   name_prefix: "all_grad_alignment_experiment_1stage_d1consd2cons"
   n_stages: 1
-  n_seeds: 10
+  n_seeds: 15
   n_seeds_stage2: 1
-  start_seed: 900
+  start_seed: 910
   slurm: True
   n_gpu_hours: 36
 

configs/grad_alignment/new_experiments/pythia1B_bs16384_2stage_qdconsincons.yaml

@@ -0,0 +1,83 @@
+data_arguments:
+  dataset: "cvdb"
+  block_size: 48
+  label_block_size: 8
+  train_subset: 'full'
+  num_ents: 4000
+
+  frac_n_qd1consis: 0.25
+  frac_n_qd1incons: 0.0
+  frac_n_qd2incons: 0.25
+  frac_n_q: 0.1
+  frac_n_d1consis: 0.1
+  frac_n_d2consis: 0.1
+  frac_n_no_qd_baseline: 0.08
+  frac_n_q_no_replacement_baseline: 0.12
+
+
+model_arguments:
+  seq2seq: False
+  max_new_tokens: 8
+  model_name_or_path: "EleutherAI/pythia-1b-deduped"
+  #model_name_or_path: "EleutherAI/pythia-410m-deduped"
+  # model_name_or_path: "gpt2"
+  # model_name_or_path: "EleutherAI/pythia-160m-deduped"
+
+
+training_arguments:
+  output_dir: 'experiments/temp'
+  bf16: True
+  per_device_train_batch_size: 256
+  per_device_eval_batch_size: 256
+  num_train_epochs: 10
+  optim: "adafactor"
+  overwrite_output_dir: True
+  auto_find_batch_size: False
+  save_strategy: "no"
+  load_best_model_at_end: False
+  evaluation_strategy: 'epoch'
+
+  do_train: True
+  do_eval: True
+  do_sweeps: False
+  save_each_epochs: 0
+  eval_each_epochs: 15
+  eval_callback_type: "pipeline"  # pipeline or generate
+  calculate_grad_variance: True
+  grad_keys: "train_defs_qd1consis,train_defs_qd2incons,qd1consis,qd2incons"
+
+experiment_arguments: # common experiment arguments
+  define_experiment: True
+  numeric_experiment: False
+  name_prefix: "all_grad_alignment_experiment_1stage_qd1consqd2incons"
+  n_stages: 1
+  n_seeds: 5
+  n_seeds_stage2: 1
+  start_seed: 900
+  slurm: True
+  n_gpu_hours: 36
+
+
+define_experiment_arguments:
+  def_order: "tve"
+  entity_association_test_sets: True
+
+
+numeric_experiment_arguments:
+  modular_experiment_baseline: False
+  modular_experiment: False
+  num_choice_experiment: False
+
+
+# overrides specified parameters
+first_stage_arguments:
+  train_subset: 'stage1'
+  num_train_epochs: 151
+  gradient_accumulation_steps: 64
+
+second_stage_arguments:
+  train_subset: 'stage2'
+  num_train_epochs: 10
+  gradient_accumulation_steps: 16
+  dont_save_in_the_end: True
+  save_each_epochs: 0

configs/grad_alignment/new_experiments/pythia1B_bs2048_1stage_d1consd2cons.yaml

@@ -0,0 +1,83 @@
+data_arguments:
+  dataset: "cvdb"
+  block_size: 48
+  label_block_size: 8
+  train_subset: 'full'
+  num_ents: 4000
+
+  frac_n_qd1consis: 0.25
+  frac_n_qd1incons: 0.0
+  frac_n_qd2incons: 0.25
+  frac_n_q: 0.1
+  frac_n_d1consis: 0.1
+  frac_n_d2consis: 0.1
+  frac_n_no_qd_baseline: 0.08
+  frac_n_q_no_replacement_baseline: 0.12
+
+
+model_arguments:
+  seq2seq: False
+  max_new_tokens: 8
+  model_name_or_path: "EleutherAI/pythia-1b-deduped"
+  #model_name_or_path: "EleutherAI/pythia-410m-deduped"
+  # model_name_or_path: "gpt2"
+  # model_name_or_path: "EleutherAI/pythia-160m-deduped"
+
+
+training_arguments:
+  output_dir: 'experiments/temp'
+  bf16: True
+  per_device_train_batch_size: 256
+  per_device_eval_batch_size: 256
+  num_train_epochs: 10
+  optim: "adafactor"
+  overwrite_output_dir: True
+  auto_find_batch_size: False
+  save_strategy: "no"
+  load_best_model_at_end: False
+  evaluation_strategy: 'epoch'
+
+  do_train: True
+  do_eval: True
+  do_sweeps: False
+  save_each_epochs: 0
+  eval_each_epochs: 2
+  eval_callback_type: "pipeline"  # pipeline or generate
+  calculate_grad_variance: True
+  grad_keys: "train_defs_d1consis,train_defs_d2consis,d1consis,d2consis"
+
+experiment_arguments: # common experiment arguments
+  define_experiment: True
+  numeric_experiment: False
+  name_prefix: "all_grad_alignment_experiment_1stage_d1consd2cons"
+  n_stages: 1
+  n_seeds: 10
+  n_seeds_stage2: 1
+  start_seed: 900
+  slurm: True
+  n_gpu_hours: 36
+
+
+define_experiment_arguments:
+  def_order: "tve"
+  entity_association_test_sets: True
+
+
+numeric_experiment_arguments:
+  modular_experiment_baseline: False
+  modular_experiment: False
+  num_choice_experiment: False
+
+
+# overrides specified parameters
+first_stage_arguments:
+  train_subset: 'stage1'
+  num_train_epochs: 41
+  gradient_accumulation_steps: 8
+
+second_stage_arguments:
+  train_subset: 'stage2'
+  num_train_epochs: 10
+  gradient_accumulation_steps: 8
+  dont_save_in_the_end: True
+  save_each_epochs: 0

configs/grad_alignment/new_experiments/pythia1B_bs512_1stage_d1consd2cons.yaml

@@ -0,0 +1,83 @@
+data_arguments:
+  dataset: "cvdb"
+  block_size: 48
+  label_block_size: 8
+  train_subset: 'full'
+  num_ents: 4000
+
+  frac_n_qd1consis: 0.25
+  frac_n_qd1incons: 0.0
+  frac_n_qd2incons: 0.25
+  frac_n_q: 0.1
+  frac_n_d1consis: 0.1
+  frac_n_d2consis: 0.1
+  frac_n_no_qd_baseline: 0.08
+  frac_n_q_no_replacement_baseline: 0.12
+
+
+model_arguments:
+  seq2seq: False
+  max_new_tokens: 8
+  model_name_or_path: "EleutherAI/pythia-1b-deduped"
+  #model_name_or_path: "EleutherAI/pythia-410m-deduped"
+  # model_name_or_path: "gpt2"
+  # model_name_or_path: "EleutherAI/pythia-160m-deduped"
+
+
+training_arguments:
+  output_dir: 'experiments/temp'
+  bf16: True
+  per_device_train_batch_size: 256
+  per_device_eval_batch_size: 256
+  num_train_epochs: 10
+  optim: "adafactor"
+  overwrite_output_dir: True
+  auto_find_batch_size: False
+  save_strategy: "no"
+  load_best_model_at_end: False
+  evaluation_strategy: 'epoch'
+
+  do_train: True
+  do_eval: True
+  do_sweeps: False
+  save_each_epochs: 0
+  eval_each_epochs: 2
+  eval_callback_type: "pipeline"  # pipeline or generate
+  calculate_grad_variance: True
+  grad_keys: "train_defs_d1consis,train_defs_d2consis,d1consis,d2consis"
+
+experiment_arguments: # common experiment arguments
+  define_experiment: True
+  numeric_experiment: False
+  name_prefix: "all_grad_alignment_experiment_1stage_d1consd2cons"
+  n_stages: 1
+  n_seeds: 10
+  n_seeds_stage2: 1
+  start_seed: 900
+  slurm: True
+  n_gpu_hours: 36
+
+
+define_experiment_arguments:
+  def_order: "tve"
+  entity_association_test_sets: True
+
+
+numeric_experiment_arguments:
+  modular_experiment_baseline: False
+  modular_experiment: False
+  num_choice_experiment: False
+
+
+# overrides specified parameters
+first_stage_arguments:
+  train_subset: 'stage1'
+  num_train_epochs: 21
+  gradient_accumulation_steps: 2
+
+second_stage_arguments:
+  train_subset: 'stage2'
+  num_train_epochs: 10
+  gradient_accumulation_steps: 2
+  dont_save_in_the_end: True
+  save_each_epochs: 0

configs/grad_alignment/new_experiments/pythia1B_bs8k_1stage_d1consd2cons.yaml

@@ -0,0 +1,83 @@
+data_arguments:
+  dataset: "cvdb"
+  block_size: 48
+  label_block_size: 8
+  train_subset: 'full'
+  num_ents: 4000
+
+  frac_n_qd1consis: 0.25
+  frac_n_qd1incons: 0.0
+  frac_n_qd2incons: 0.25
+  frac_n_q: 0.1
+  frac_n_d1consis: 0.1
+  frac_n_d2consis: 0.1
+  frac_n_no_qd_baseline: 0.08
+  frac_n_q_no_replacement_baseline: 0.12
+
+
+model_arguments:
+  seq2seq: False
+  max_new_tokens: 8
+  model_name_or_path: "EleutherAI/pythia-1b-deduped"
+  #model_name_or_path: "EleutherAI/pythia-410m-deduped"
+  # model_name_or_path: "gpt2"
+  # model_name_or_path: "EleutherAI/pythia-160m-deduped"
+
+
+training_arguments:
+  output_dir: 'experiments/temp'
+  bf16: True
+  per_device_train_batch_size: 256
+  per_device_eval_batch_size: 256
+  num_train_epochs: 10
+  optim: "adafactor"
+  overwrite_output_dir: True
+  auto_find_batch_size: False
+  save_strategy: "no"
+  load_best_model_at_end: False
+  evaluation_strategy: 'epoch'
+
+  do_train: True
+  do_eval: True
+  do_sweeps: False
+  save_each_epochs: 0
+  eval_each_epochs: 5
+  eval_callback_type: "pipeline"  # pipeline or generate
+  calculate_grad_variance: True
+  grad_keys: "train_defs_d1consis,train_defs_d2consis,d1consis,d2consis"
+
+experiment_arguments: # common experiment arguments
+  define_experiment: True
+  numeric_experiment: False
+  name_prefix: "all_grad_alignment_experiment_1stage_d1consd2cons"
+  n_stages: 1
+  n_seeds: 10
+  n_seeds_stage2: 1
+  start_seed: 900
+  slurm: True
+  n_gpu_hours: 36
+
+
+define_experiment_arguments:
+  def_order: "tve"
+  entity_association_test_sets: True
+
+
+numeric_experiment_arguments:
+  modular_experiment_baseline: False
+  modular_experiment: False
+  num_choice_experiment: False
+
+
+# overrides specified parameters
+first_stage_arguments:
+  train_subset: 'stage1'
+  num_train_epochs: 81
+  gradient_accumulation_steps: 32
+
+second_stage_arguments:
+  train_subset: 'stage2'
+  num_train_epochs: 10
+  gradient_accumulation_steps: 32
+  dont_save_in_the_end: True
+  save_each_epochs: 0