43 lines
862 B
YAML
43 lines
862 B
YAML
# 之之母模型 v2.0 · 全参数 SFT
|
|
# 由铸渊配置 · 2026-05-18
|
|
# 基座: Qwen2.5-7B · 数据: 铸渊仓库 + 霜砚 混合语料
|
|
# 数据量: 1868 条
|
|
|
|
### 模型 ###
|
|
model_name_or_path: Qwen/Qwen2.5-7B
|
|
|
|
### 数据 ###
|
|
dataset: zhizhi_v2
|
|
dataset_dir: /root/autodl-tmp
|
|
template: qwen
|
|
cutoff_len: 4096
|
|
|
|
### 训练 ###
|
|
finetuning_type: full
|
|
per_device_train_batch_size: 1
|
|
gradient_accumulation_steps: 8
|
|
learning_rate: 5.0e-6
|
|
num_train_epochs: 3.0
|
|
lr_scheduler_type: cosine
|
|
warmup_ratio: 0.03
|
|
|
|
### 精度 ###
|
|
optim: adamw_torch
|
|
bf16: true
|
|
mixed_precision: bf16
|
|
ddp_find_unused_parameters: false
|
|
|
|
### DeepSpeed ###
|
|
deepspeed: /root/autodl-tmp/ds_config.json
|
|
|
|
### 输出 ###
|
|
output_dir: /root/autodl-tmp/output/zhizhi_v2
|
|
logging_steps: 5
|
|
save_steps: 58
|
|
plot_loss: true
|
|
overwrite_output_dir: true
|
|
preprocessing_num_workers: 16
|
|
|
|
### 报告 ###
|
|
report_to: none
|