liqi / code /pmps /args /gsm_codi_direct_match_train.yaml

Upload liqi files

0ab2566 verified 1 day ago

2.55 kB

# Ablation: Direct Soft Matching (no prototypes)
# Based on gsm_codi_pmps_train.yaml — only the model class changes (via train_direct_match.py)

model_name_or_path: /data/oss_bucket_0/liqi/models/gpt2
lora_r: 128
lora_alpha: 32
lora_init: True
use_lora: True
do_train: True

# ==================== Data Arguments ====================
data_name: icot

# ==================== Training Arguments ====================
output_dir: /data/oss_bucket_0/liqi/codi_pmps/ablation_direct_match_mt0.1_fac0.5_l6_lr3e3_btz256_pos2_ep30
expt_name: gsm8k_gpt_direct_match
logging_dir: /data/oss_bucket_0/liqi/codi_pmps/ablation_direct_match_mt0.1_fac0.5_l6_lr3e3_btz256_pos2_ep30/logs
logging_steps: 10
logging_strategy: steps

enable_top_k: False
latent_top_k: 0
latent_top_k_temperature: 1.0

# ===== Direct Match Loss (uses same yaml keys for compatibility) =====
# DirectMatchSupervisor uses: pmps_loss_factor, pmps_match_temperature,
# pmps_feat_dim, pmps_hidden_mlp, pmps_use_projector, pmps_layer
# Other PMPS keys (temperature, epsilon, sinkhorn_iterations, nmb_prototypes)
# are parsed but unused by DirectMatchSupervisor.
pmps_loss_factor: 0.5
pmps_temperature: 0.1
pmps_match_temperature: 0.1
pmps_epsilon: 0.05
pmps_sinkhorn_iterations: 3
pmps_nmb_prototypes: 24
pmps_feat_dim: 128
pmps_hidden_mlp: 512
pmps_use_projector: True
pmps_freeze_prototypes_steps: 0
pmps_layer: last
pmps_start_epoch: 0

# ===== Progressive Sequential Alignment (PSA) =====
pmps_position_bias_max: 2.0
pmps_position_bias_schedule: constant

# ===== Latent Diversity Regularisation =====
# Recommended > 0 for direct matching (no Sinkhorn to prevent collapse)
diversity_loss_factor: 0.0

seed: 42
model_max_length: 512
per_device_train_batch_size: 32
gradient_accumulation_steps: 1
bf16: True
num_train_epochs: 30
learning_rate: !!float 3e-3
max_grad_norm: 2.0

save_strategy: "epoch"
save_safetensors: False
save_total_limit: 50
weight_decay: 0.1
warmup_ratio: 0.03
lr_scheduler_type: cosine
report_to: wandb
include_last_cot: False

# ==================== Wandb Arguments ====================
use_wandb: True
wandb_api_key: "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"
wandb_project: codi-pmps
wandb_run_name: ablation_direct_match_mt0.1_fac0.5_l6_lr3e3_btz256_pos2_ep30

# ==================== CODI Specific Arguments ====================
num_latent: 6
use_prj: True
prj_dim: 768
prj_dropout: 0.0
distill_loss_div_std: True
exp_mode: False
exp_data_num: 2000
remove_eos: True
print_ref_model_stats: True
distill_loss_factor: 1.0

	# Ablation: Direct Soft Matching (no prototypes)
	# Based on gsm_codi_pmps_train.yaml — only the model class changes (via train_direct_match.py)

	model_name_or_path: /data/oss_bucket_0/liqi/models/gpt2
	lora_r: 128
	lora_alpha: 32
	lora_init: True
	use_lora: True
	do_train: True

	# ==================== Data Arguments ====================
	data_name: icot

	# ==================== Training Arguments ====================
	output_dir: /data/oss_bucket_0/liqi/codi_pmps/ablation_direct_match_mt0.1_fac0.5_l6_lr3e3_btz256_pos2_ep30
	expt_name: gsm8k_gpt_direct_match
	logging_dir: /data/oss_bucket_0/liqi/codi_pmps/ablation_direct_match_mt0.1_fac0.5_l6_lr3e3_btz256_pos2_ep30/logs
	logging_steps: 10
	logging_strategy: steps

	enable_top_k: False
	latent_top_k: 0
	latent_top_k_temperature: 1.0

	# ===== Direct Match Loss (uses same yaml keys for compatibility) =====
	# DirectMatchSupervisor uses: pmps_loss_factor, pmps_match_temperature,
	# pmps_feat_dim, pmps_hidden_mlp, pmps_use_projector, pmps_layer
	# Other PMPS keys (temperature, epsilon, sinkhorn_iterations, nmb_prototypes)
	# are parsed but unused by DirectMatchSupervisor.
	pmps_loss_factor: 0.5
	pmps_temperature: 0.1
	pmps_match_temperature: 0.1
	pmps_epsilon: 0.05
	pmps_sinkhorn_iterations: 3
	pmps_nmb_prototypes: 24
	pmps_feat_dim: 128
	pmps_hidden_mlp: 512
	pmps_use_projector: True
	pmps_freeze_prototypes_steps: 0
	pmps_layer: last
	pmps_start_epoch: 0

	# ===== Progressive Sequential Alignment (PSA) =====
	pmps_position_bias_max: 2.0
	pmps_position_bias_schedule: constant

	# ===== Latent Diversity Regularisation =====
	# Recommended > 0 for direct matching (no Sinkhorn to prevent collapse)
	diversity_loss_factor: 0.0

	seed: 42
	model_max_length: 512
	per_device_train_batch_size: 32
	gradient_accumulation_steps: 1
	bf16: True
	num_train_epochs: 30
	learning_rate: !!float 3e-3
	max_grad_norm: 2.0

	save_strategy: "epoch"
	save_safetensors: False
	save_total_limit: 50
	weight_decay: 0.1
	warmup_ratio: 0.03
	lr_scheduler_type: cosine
	report_to: wandb
	include_last_cot: False

	# ==================== Wandb Arguments ====================
	use_wandb: True
	wandb_api_key: "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"
	wandb_project: codi-pmps
	wandb_run_name: ablation_direct_match_mt0.1_fac0.5_l6_lr3e3_btz256_pos2_ep30

	# ==================== CODI Specific Arguments ====================
	num_latent: 6
	use_prj: True
	prj_dim: 768
	prj_dropout: 0.0
	distill_loss_div_std: True
	exp_mode: False
	exp_data_num: 2000
	remove_eos: True
	print_ref_model_stats: True
	distill_loss_factor: 1.0