---
library_name: transformers
tags:
- generated_from_trainer
---

<!-- 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"/>](https://github.com/axolotl-ai-cloud/axolotl)
<details><summary>See axolotl config</summary>

axolotl version: `0.16.0.dev0`
```yaml
# config-4gpu-fullft-e4b-32k.yml
base_model: /models/gemma-4-e4b-it

embeddings_skip_upcast: true
trust_remote_code: true
chat_template: gemma
unfrozen_parameters:
  - model.language_model.layers.(2|3|4)[\d].(_checkpoint_wrapped_module.)?(mlp).(up|down|gate)_proj
# ====================== 多 GPU 設定 (FSDP) ======================
fsdp_version: 2
fsdp_config:
  offload_params: false
  state_dict_type: FULL_STATE_DICT
  auto_wrap_policy: TRANSFORMER_BASED_WRAP
  transformer_layer_cls_to_wrap: Gemma4TextDecoderLayer
# ====================== Liger Kernel ======================
plugins:
  - axolotl.integrations.liger.LigerPlugin

torch_compile: false
liger_layer_norm: false
liger_rope: true
liger_rms_norm: true
liger_glu_activation: true
liger_rms_norm_gated: true
sdp_attention: true

# ====================== 資料集 ======================
datasets:
  - path: /notebook/train_segments.jsonl
    type: input_output

dataset_processes: 4
sample_packing: true
pad_to_sequence_len: true
eval_sample_packing: false

# ====================== 關鍵：長上下文 32768 ======================
sequence_len: 16384
micro_batch_size: 1 # 32k 必須從 1 開始，避免 OOM
gradient_accumulation_steps: 1 # effective batch size ≈ 1×4×8 = 32（推薦 DPO 值）
max_grad_norm: 1
num_epochs: 2

# 記憶體優化（32k 長上下文非常吃 activations）
gradient_checkpointing: true
activation_offloading: false # 強烈建議開啟

# 優化器
optimizer: adamw_torch
lr_scheduler: constant
learning_rate: 5e-6

# 混合精度
bf16: true
tf32: true

# 保存與紀錄
save_safetensors: true
save_strategy: epoch
saves_per_epoch: 1
logging_steps: 5 # 長上下文時 logging 頻率提高一點
output_dir: ./outputs/gemma4-e4b-sft-4gpu-fullft-32k

use_tensorboard: true
#hub_model_id: AlexHung29629/WhiteDubstepFly

```

</details><br>

# outputs/gemma4-e4b-sft-4gpu-fullft-32k

This model was trained from scratch on the /notebook/train_segments.jsonl dataset.

## 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: 5e-06
- train_batch_size: 1
- eval_batch_size: 1
- seed: 42
- distributed_type: multi-GPU
- num_devices: 4
- total_train_batch_size: 4
- total_eval_batch_size: 4
- optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
- lr_scheduler_type: constant
- lr_scheduler_warmup_steps: 7
- training_steps: 262

### Training results



### Framework versions

- Transformers 5.5.0
- Pytorch 2.10.0+cu130
- Datasets 4.5.0
- Tokenizers 0.22.2