# CODI Testing Configuration for nebulactl run mdl
# This config file is designed to work with the test_nebula.py wrapper

# ==================== Model Arguments ====================
model_name_or_path: /data/oss_bucket_0/liqi/models/gpt2
lora_r: 128
lora_alpha: 32
lora_init: True
use_lora: True
# ckpt_dir: /data/oss_bucket_0/liqi/models/SIM_COT-GPT2-CODI
ckpt_dir: /data/oss_bucket_0/liqi/codi_pmps/ablation_direct_match_mt0.1_fac0.5_l6_lr3e3_btz256_pos2_div0.1_ep50/gsm8k_gpt_direct_match/
# ckpt_dir: /data/oss_bucket_0/liqi/codi_pmps/codi_ori_single/gsm8k_gpt_latent_decoder/

# ==================== Data Arguments ====================
# data_name: gsm8k
# data_name: gsm-hard
# data_name: multi-arith
data_name: svamp
batch_size: 128

enable_top_k: False
latent_top_k: 20
latent_top_k_temperature: 1.0

# ==================== Training Arguments (for inference) ====================
# output_dir: /data/oss_bucket_0/liqi/llm_baseline/sim-cot-official/gpt2/output_svamp_new
output_dir: /data/oss_bucket_0/liqi/codi_pmps/ablation_direct_match_mt0.1_fac0.5_l6_lr3e3_btz256_pos2_div0.1_ep50/output-svamp
model_max_length: 512
bf16: True

# ==================== CODI Specific Arguments ====================
pmps_nmb_prototypes: 24
use_prj: True
prj_dim: 768
prj_dropout: 0.0
prj_no_ln: False
inf_latent_iterations: 6
inf_num_iterations: 1
remove_eos: True
greedy: True