cjkasbdkjnlakb/agent-0918-xml-5k
010
<!-- This model card has been generated automatically according to the information the Trainer had access to. You should probably proofread and complete it, then remove this comment. -->
<img src="https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/> <details><summary>See axolotl config</summary>
axolotl version: 0.12.2
# Automatically upload checkpoint and final model to HF
# hub_model_id: username/custom_model_name
# 是否以 8-bit 精度加载模型
load_in_8bit: false
# 是否以 4-bit 精度加载模型(与QLoRA绑定, 强制使用)
load_in_4bit: false
# 是否严格匹配模型结构,关闭表示可加载少部分差异结构(如以适配 adapter)
# strict: false
base_model: Qwen/Qwen3-4B-Instruct-2507
# 数据集设置
chat_template: qwen3
datasets:
- path: /workspace/train_dir/tool_and_retrieval_agent_train_data_xml_5k.json # - 表示列表(list)中的一项, 即可以同时使用多个数据集
type: chat_template # chat_template(自定义格式) alpaca
roles_to_train: ["assistant"]
field_messages: messages # 标识的字段
message_property_mappings: # message_property_mappings={'role':'role', 'content':'content'})
role: role
content: content
dataset_prepared_path:
val_set_size: 0.05
output_dir: checkpoints/0918-xml-5k
sequence_len: 16384 # 模型所能处理的最大上下文长度(默认2048)
pad_to_sequence_len: true
# context_parallel_size: 2 # 长序列拆分至多个GPU(强制要求 mirco_batch_size: 1)
sample_packing: false # 在训练时将多个样本拼接(packing)成一个长序列(sequence_len)输入到模型中,以提高训练效率。
eval_sample_packing: false # 评估时拼接多个样本
# 训练超参数
adapter: lora # lora qlora
lora_model_dir:
lora_r: 16 # lora_r默认首选 16,平衡精度与显存
lora_alpha: 64 # 缩放系数,用于控制 LoRA 的影响力, 一般设为 2*r 或 4*r
lora_dropout: 0.05
lora_target_linear: true
micro_batch_size: 4 # 微批次大小 94G的H100可以设为4(Token为1w)
gradient_accumulation_steps: 2 # 梯度累积: 将多个微批次的梯度(micro_batch_size)累积起来,然后更新模型权重 有效 Batch 常取 16: 小于 8 训练会抖,大于 32 只会更耗时、收益有限
auto_find_batch_size: false # 允许Axolotl不断调整batch_size ⚠️Zero-3不适用
num_epochs: 1
optimizer: adamw_torch_fused
lr_scheduler: cosine
learning_rate: 4e-5
# bf16: auto + tf32: true,可获得更好的稳定性和性能。
bf16: auto
tf32: true
# early_stopping_patience:
gradient_checkpointing: true
gradient_checkpointing_kwargs:
use_reentrant: false
# auto_resume_from_checkpoints: true #自动从output_dir寻找最新checkpoint断点恢复
logging_steps: 1
flash_attention: true
warmup_steps: 50
evals_per_epoch: 4
saves_per_epoch: 1
weight_decay: 0.01
fsdp:
- full_shard
- auto_wrap
fsdp_config:
fsdp_limit_all_gathers: true
fsdp_sync_module_states: true
fsdp_offload_params: false # H200显存足够,无需offload
fsdp_use_orig_params: false
fsdp_cpu_ram_efficient_loading: true
fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
fsdp_transformer_layer_cls_to_wrap: Qwen3DecoderLayer
fsdp_state_dict_type: FULL_STATE_DICT
fsdp_sharding_strategy: FULL_SHARD</details><br>
checkpoints/0918-xml-5k
This model is a fine-tuned version of Qwen/Qwen3-4B-Instruct-2507 on the /workspace/traindir/toolandretrievalagenttraindataxml5k.json dataset. It achieves the following results on the evaluation set:
- Loss: 0.0774
- Memory/max Mem Active(gib): 128.99
- Memory/max Mem Allocated(gib): 128.8
- Memory/device Mem Reserved(gib): 130.32
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
- trainbatchsize: 4
- evalbatchsize: 4
- seed: 42
- distributed_type: multi-GPU
- num_devices: 4
- gradientaccumulationsteps: 2
- totaltrainbatch_size: 32
- totalevalbatch_size: 16
- optimizer: Use OptimizerNames.ADAMWTORCHFUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
- lrschedulertype: cosine
- lrschedulerwarmup_steps: 50
- training_steps: 149
Training results
Framework versions
- PEFT 0.17.0
- Transformers 4.55.2
- Pytorch 2.6.0+cu126
- Datasets 4.0.0
- Tokenizers 0.21.4
