liqi / code /pmps /args /analysis /analysis_codi_gpt2.yaml
WaitXian's picture
Upload liqi files
0ab2566 verified
Raw History Blame Contribute Delete
1.45 kB
# ================================================================
# CODI (baseline) 分析实验配置 (GPT-2, GSM8K)
# 用于 extract_latent_states.py(不需要 --extract_pmps)
# ================================================================
#
# CODI = PMPS with pmps_loss_factor=0,所以模型结构中不会创建 pmps_module。
# 但 LoRA / projection / num_latent 等结构参数必须和 CODI 训练时一致。
# ==================== 必填:模型和 Checkpoint ====================
model_name_or_path: /data/oss_bucket_0/liqi/models/gpt2
ckpt_dir: <填写你的 CODI checkpoint 路径>
# ==================== 必须和训练一致的参数 ====================
lora_r: 128
lora_alpha: 32
lora_init: True
use_lora: True
num_latent: 6
use_prj: True
prj_dim: 768
# --- 关键区别:pmps_loss_factor = 0 → 没有 pmps_module ---
pmps_loss_factor: 0.0
pmps_nmb_prototypes: 24 # 虽然 factor=0 不会创建 module,但 parser 需要
pmps_feat_dim: 128
pmps_hidden_mlp: 512
pmps_use_projector: True
pmps_temperature: 0.1
pmps_match_temperature: 0.3
pmps_epsilon: 0.05
pmps_sinkhorn_iterations: 3
pmps_layer: last
pmps_position_bias_max: 0.0
pmps_position_bias_schedule: cosine
enable_top_k: False
latent_top_k: 0
latent_top_k_temperature: 1.0
remove_eos: True
diversity_loss_factor: 0.0
# ==================== 其他 ====================
data_name: gsm8k
model_max_length: 512
bf16: True
output_dir: ./analysis/results/codi