liqi / code /pmps /args /gsm_codi_pmps_llama_train_math.yaml

Upload liqi files

0ab2566 verified 4 days ago

2.71 kB

# CODI + PMPS Training Configuration for GSM8k
# Based on codi_ori/args/gsm_codi_train.yaml with PMPS-specific additions

model_name_or_path: /data/oss_bucket_0/liqi/models/Llama-3.2-3B-Instruct
# model_name_or_path: /data/oss_bucket_0/liqi/models/Qwen3-4B-Instruct-2507
lora_r: 128
lora_alpha: 32
lora_init: True
use_lora: True
do_train: True

# ==================== Data Arguments ====================
data_name: math


# ==================== Training Arguments ====================

output_dir: /data/oss_bucket_0/liqi/larger_model/llama-3b/math/k32_t0.1_mt0.1_fac0.3_l8_lr9e4_btz256_pos6_ep300
logging_dir: /data/oss_bucket_0/liqi/larger_model/llama-3b/math/k32_t0.1_mt0.1_fac0.3_l8_lr9e4_btz256_pos6_ep300/logs
expt_name: gsm8k_gpt_pmps
logging_steps: 10
logging_strategy: steps

enable_top_k: False
latent_top_k: 30
latent_top_k_temperature: 1.0

# ===== PMPS Loss (NEW) =====
pmps_loss_factor: 0.3
pmps_temperature: 0.1
pmps_match_temperature: 0.1
pmps_epsilon: 0.05
pmps_sinkhorn_iterations: 3
pmps_nmb_prototypes: 32     # ~4 × num_latent (num_latent=6 -> K=24)
pmps_feat_dim: 128
pmps_hidden_mlp: 512
# pmps_feat_dim: 256
# pmps_hidden_mlp: 1024
pmps_use_projector: True
pmps_freeze_prototypes_steps: 0
pmps_layer: last
pmps_start_epoch: 0

# ===== Progressive Sequential PMPS (NEW) =====
pmps_position_bias_max: 6.0       # Max positional bias strength. 0=disabled. Recommended: 3.0~8.0
pmps_position_bias_schedule: cosine   # Decay schedule: cosine / linear / constant

# ===== Latent Diversity Regularisation (NEW) =====
diversity_loss_factor: 0.0     # Weight for diversity loss. 0=disabled. Recommended: 0.05~0.3

seed: 11
model_max_length: 512
per_device_train_batch_size: 16
# per_device_train_batch_size: 32
# per_device_train_batch_size: 8
# gradient_accumulation_steps: 2
gradient_accumulation_steps: 2
bf16: True
num_train_epochs: 300
learning_rate: !!float 9e-4
# learning_rate: !!float 1e-3
max_grad_norm: 2.0
max_token_num: 200

save_strategy: "epoch"
save_safetensors: False
save_total_limit: 40
weight_decay: 0.1
warmup_ratio: 0.03
lr_scheduler_type: cosine
report_to: wandb
include_last_cot: True

# ==================== Wandb Arguments ====================
use_wandb: True
wandb_api_key: "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"  # Fill in your wandb API key here
wandb_project: codi-pmps-llama-3B-math
wandb_run_name: k32_t0.1_mt0.1_fac0.3_l8_lr9e4_btz256_pos6_ep300

# ==================== CODI Specific Arguments ====================
num_latent: 8
use_prj: True
prj_dim: 3072
prj_dropout: 0.0
distill_loss_div_std: True
exp_mode: False
exp_data_num: 200
distill_loss_factor: 20.0
remove_eos: True
print_ref_model_stats: True

	# CODI + PMPS Training Configuration for GSM8k
	# Based on codi_ori/args/gsm_codi_train.yaml with PMPS-specific additions

	model_name_or_path: /data/oss_bucket_0/liqi/models/Llama-3.2-3B-Instruct
	# model_name_or_path: /data/oss_bucket_0/liqi/models/Qwen3-4B-Instruct-2507
	lora_r: 128
	lora_alpha: 32
	lora_init: True
	use_lora: True
	do_train: True

	# ==================== Data Arguments ====================
	data_name: math


	# ==================== Training Arguments ====================

	output_dir: /data/oss_bucket_0/liqi/larger_model/llama-3b/math/k32_t0.1_mt0.1_fac0.3_l8_lr9e4_btz256_pos6_ep300
	logging_dir: /data/oss_bucket_0/liqi/larger_model/llama-3b/math/k32_t0.1_mt0.1_fac0.3_l8_lr9e4_btz256_pos6_ep300/logs
	expt_name: gsm8k_gpt_pmps
	logging_steps: 10
	logging_strategy: steps

	enable_top_k: False
	latent_top_k: 30
	latent_top_k_temperature: 1.0

	# ===== PMPS Loss (NEW) =====
	pmps_loss_factor: 0.3
	pmps_temperature: 0.1
	pmps_match_temperature: 0.1
	pmps_epsilon: 0.05
	pmps_sinkhorn_iterations: 3
	pmps_nmb_prototypes: 32 # ~4 × num_latent (num_latent=6 -> K=24)
	pmps_feat_dim: 128
	pmps_hidden_mlp: 512
	# pmps_feat_dim: 256
	# pmps_hidden_mlp: 1024
	pmps_use_projector: True
	pmps_freeze_prototypes_steps: 0
	pmps_layer: last
	pmps_start_epoch: 0

	# ===== Progressive Sequential PMPS (NEW) =====
	pmps_position_bias_max: 6.0 # Max positional bias strength. 0=disabled. Recommended: 3.0~8.0
	pmps_position_bias_schedule: cosine # Decay schedule: cosine / linear / constant

	# ===== Latent Diversity Regularisation (NEW) =====
	diversity_loss_factor: 0.0 # Weight for diversity loss. 0=disabled. Recommended: 0.05~0.3

	seed: 11
	model_max_length: 512
	per_device_train_batch_size: 16
	# per_device_train_batch_size: 32
	# per_device_train_batch_size: 8
	# gradient_accumulation_steps: 2
	gradient_accumulation_steps: 2
	bf16: True
	num_train_epochs: 300
	learning_rate: !!float 9e-4
	# learning_rate: !!float 1e-3
	max_grad_norm: 2.0
	max_token_num: 200

	save_strategy: "epoch"
	save_safetensors: False
	save_total_limit: 40
	weight_decay: 0.1
	warmup_ratio: 0.03
	lr_scheduler_type: cosine
	report_to: wandb
	include_last_cot: True

	# ==================== Wandb Arguments ====================
	use_wandb: True
	wandb_api_key: "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" # Fill in your wandb API key here
	wandb_project: codi-pmps-llama-3B-math
	wandb_run_name: k32_t0.1_mt0.1_fac0.3_l8_lr9e4_btz256_pos6_ep300

	# ==================== CODI Specific Arguments ====================
	num_latent: 8
	use_prj: True
	prj_dim: 3072
	prj_dropout: 0.0
	distill_loss_div_std: True
	exp_mode: False
	exp_data_num: 200
	distill_loss_factor: 20.0
	remove_eos: True
	print_ref_model_stats: True