See axolotl config
axolotl version: 0.12.2
# --- QLoRAでの最小CPT設定 ---
base_model: Qwen/Qwen3-8B
tokenizer_config: Qwen/Qwen3-8B
# プレトレ用の素朴なJSONL（各行 {"text": "..."}）
pretraining_dataset:
- path: json
data_files:
- data.jsonl
field: text
# 長さとパッキング
sequence_len: 2048
sample_packing: false
pad_to_sequence_len: false
train_on_inputs: false
# 演算系
bf16: true
flash_attention: false
attn_implementation: sdpa
# バッチ/ステップ（まずは通るか確認）
micro_batch_size: 24
gradient_accumulation_steps: 1
max_steps: 150
# 最適化
optimizer: adamw_torch
learning_rate: 1.0e-4
weight_decay: 0.1
lr_scheduler: cosine
warmup_ratio: 0.01
# ロギング/保存
logging_steps: 2
save_steps: 20
output_dir: ./ckpts/Qwen3-8B-cpt
wandb_project: null
# QLoRA（ここが重要）
adapter: lora
load_in_4bit: true
bnb_4bit_quant_type: nf4
bnb_4bit_use_double_quant: true
bnb_4bit_compute_dtype: bfloat16
# LoRAハイパラ
lora_r: 64
lora_alpha: 128
lora_dropout: 0.05
lora_target_modules:
- q_proj
- k_proj
- v_proj
- o_proj
- gate_proj
- up_proj
- down_proj
# 省メモリ/IO
gradient_checkpointing: true
dataloader_num_workers: 1
dataset_processes: 1
dataloader_prefetch_factor: 8
ckpts/Qwen3-8B-cpt
This model is a fine-tuned version of Qwen/Qwen3-8B on an unknown 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: 0.0001
- train_batch_size: 24
- eval_batch_size: 24
- seed: 42
- distributed_type: multi-GPU
- num_devices: 3
- total_train_batch_size: 72
- total_eval_batch_size: 72
- 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: cosine
- lr_scheduler_warmup_steps: 2
- training_steps: 150
Training results
Framework versions
- PEFT 0.17.0
- Transformers 4.55.2
- Pytorch 2.6.0+cu124
- Datasets 4.0.0
- Tokenizers 0.21.4
- Downloads last month
- 1