# 基础模型配置
base_model: Qwen/Qwen3-4B-Instruct-2507
load_in_8bit: false
load_in_4bit: false  # QLoRA才需要4bit

# LoRA 适配器配置 - 这是关键部分
adapter: lora  # 明确指定使用LoRA
lora_model_dir:  # 如果有预训练的LoRA权重可以在这里指定

# LoRA 具体参数
lora_r: 64
lora_alpha: 64
lora_dropout: 0.05
lora_target_modules:  # Qwen3模型的关键模块
  - q_proj
  - k_proj
  - v_proj
  - o_proj
  - gate_proj
  - up_proj
  - down_proj
lora_target_linear: true  # 自动找到所有线性层
lora_fan_in_fan_out: false

# 数据集设置
chat_template: qwen3
datasets:
  - path: /workspace/tool_data_with_prompt.json
    type: chat_template
    roles_to_train: ["assistant"]
    field_messages: messages
    message_property_mappings:
      role: role
      content: content

dataset_prepared_path:
val_set_size: 0.05
output_dir: checkpoints

# 序列长度设置
sequence_len: 10000
pad_to_sequence_len: false
sample_packing: false
eval_sample_packing: false
group_by_length: true  # 启用长度分组，提高效率

# 训练超参数
num_epochs: 3
micro_batch_size: 1  # H100显存大
gradient_accumulation_steps: 8  # 8卡LoRA不需要太大的累积
eval_batch_size: 8

# 优化器设置
optimizer: adamw_torch_fused
lr_scheduler: cosine_with_restarts
cosine_restarts: 2  # 每个epoch重启一次
learning_rate: 4e-5
warmup_ratio: 0.03
weight_decay: 0.05

# 精度设置
bf16: auto  # H100支持bf16
tf32: true
gradient_checkpointing: true  # 节省显存
flash_attention: true

# 日志和保存
logging_steps: 30
evals_per_epoch: 1
saves_per_epoch: 1
save_total_limit: 3  # 只保留最新的3个checkpoint

# 多卡训练配置 - 使用DeepSpeed而不是FSDP
deepspeed: zero2.json  # 或者直接内联配置

# 其他优化
ddp_timeout: 3600  # DDP超时设置
ddp_find_unused_parameters: false  # LoRA通常不需要

checkpoints

This model is a fine-tuned version of Qwen/Qwen3-4B-Instruct-2507 on the /workspace/tool_data_with_prompt.json dataset. It achieves the following results on the evaluation set:

Loss: 0.0436
Memory/max Mem Active(gib): 123.28
Memory/max Mem Allocated(gib): 123.28
Memory/device Mem Reserved(gib): 124.61

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 4e-05
train_batch_size: 1
eval_batch_size: 8
seed: 42
distributed_type: multi-GPU
num_devices: 8
gradient_accumulation_steps: 8
total_train_batch_size: 64
total_eval_batch_size: 64
optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
lr_scheduler_type: cosine_with_restarts
lr_scheduler_warmup_steps: 78
training_steps: 2630

Training results

Training Loss	Epoch	Step	Validation Loss	Mem Active(gib)	Mem Allocated(gib)	Mem Reserved(gib)
No log	0	0	1.1699	123.25	123.25	124.05
0.0449	1.0	877	0.0458	123.28	123.28	124.61
0.04	2.0	1754	0.0436	123.28	123.28	124.61

Framework versions

PEFT 0.17.0
Transformers 4.55.2
Pytorch 2.6.0+cu126
Datasets 4.0.0
Tokenizers 0.21.4

cjkasbdkjnlakb/agent-1011-01

checkpoints

Model description

Intended uses & limitations

Training and evaluation data

Training procedure

Training hyperparameters

Training results

Framework versions

cjkasbdkjnlakb/
agent-1011-01