config.yaml

# Model Configuration
model:
  name: "ZhipuAI/chatglm3-6b"  # 可以换成 "bert-base-chinese" 或其他中文模型  huggingface上是zai-org/chatglm3-6b
  dir: "/root/autodl-tmp/modelscope/hub/models/ZhipuAI/chatglm3-6b"
  max_length: 128
  hidden_size: 768
  num_attention_heads: 12
  num_hidden_layers: 12
  vocab_size: 21128
  type_vocab_size: 2
  
# ✅ LoRA 配置
lora:
  r: 8  # LoRA 秩，越小参数越少（建议 4-16）
  lora_alpha: 32  # LoRA alpha，通常是 r 的 2-4 倍
  lora_dropout: 0.1  # Dropout 概率
  target_modules: ["query_key_value"]  # ChatGLM 的注意力层
  lora_path: "checkpoints/best_model"
  # 或者对于其他模型: ["q_proj", "v_proj", "k_proj", "o_proj"]

# Training Configuration  
training:
  batch_size: 8
  gradient_accumulation_steps: 4
  num_epochs: 3
  learning_rate: 0.00003
  warmup_ratio: 0.05
  weight_decay: 0.01
  max_grad_norm: 1.0
  fp16: False
  patience: 2  # 早停耐心值
  save_epochs: 1  # 每 N 个 epoch 保存一次
  gradient_checkpointing: false 
  logging_steps: 100
  save_steps: 5000
  eval_steps: 1000
  seed: 42
  
# Generation Configuration
generation:
  max_length: 50
  min_length: 5
  temperature: 0.9
  top_k: 50
  top_p: 0.95
  repetition_penalty: 1.2
  num_beams: 3
  do_sample: true
  
# Data Configuration
data:
  train_file: "data/train1.json"
  val_file: "data/val1.json"
  test_file: "data/test1.json"
  cache_dir: "data/cache"
  num_workers: 8  # ✅ 从4增加到16（关键优化，提速30%+）
  prefetch_factor: 4  # ✅ 新增：预加载更多batch
  
# Paths
paths:
  output_dir: "outputs"
  checkpoint_dir: "checkpoints"
  log_dir: "logs"
  dirty_words_file: "data/dirty_words.txt"

use_wandb: false  # 是否使用 wandb 记录