# ================================================================
# PMPS 分析实验配置 (LLaMA-3.2-1B, GSM8K)
# ================================================================

model_name_or_path: /data/oss_bucket_0/liqi/models/Llama-3.2-1B-Instruct
ckpt_dir: <填写你的 PMPS-LLaMA checkpoint 路径>

lora_r: 128
lora_alpha: 32
lora_init: True
use_lora: True

num_latent: 8
use_prj: True
prj_dim: 2048              # LLaMA-1B hidden_size=2048

pmps_loss_factor: 0.3
pmps_nmb_prototypes: 32
pmps_feat_dim: 256
pmps_hidden_mlp: 1024
pmps_use_projector: True
pmps_temperature: 0.1
pmps_match_temperature: 0.1
pmps_epsilon: 0.05
pmps_sinkhorn_iterations: 3
pmps_layer: last
pmps_position_bias_max: 6.0
pmps_position_bias_schedule: cosine

enable_top_k: False
latent_top_k: 60
latent_top_k_temperature: 1.0
remove_eos: True
diversity_loss_factor: 0.0

data_name: gsm8k
model_max_length: 512
bf16: True
output_dir: ./analysis/results/pmps_llama