mirror of
https://github.com/zenlm/zen-family.git
synced 2026-11-02 17:32:06 +00:00
47 lines
1.3 KiB
Bash
Executable File
47 lines
1.3 KiB
Bash
Executable File
#!/bin/bash
|
|
# Train zen-nano-thinking-v1.0.1 with recursive improvements
|
|
|
|
cd ~/work/zoo/gym
|
|
|
|
# Register dataset
|
|
python -c "
|
|
import json
|
|
info = json.load(open('data/dataset_info.json'))
|
|
info['zen-nano-thinking-v1.0.1_training'] = {
|
|
'file_name': 'zen-nano-thinking-v1.0.1_training.json',
|
|
'formatting': 'alpaca'
|
|
}
|
|
json.dump(info, open('data/dataset_info.json', 'w'), indent=2)
|
|
"
|
|
|
|
# Copy training data
|
|
cp ../../zen/data/zen-nano-thinking-v1.0.1_training.json data/
|
|
|
|
# Train with improvements
|
|
python src/train.py \
|
|
--stage sft \
|
|
--model_name_or_path zenlm/zen-nano-instruct \
|
|
--dataset zen-nano-thinking-v1.0.1_training \
|
|
--template qwen \
|
|
--finetuning_type lora \
|
|
--lora_target all \
|
|
--lora_rank 8 \
|
|
--lora_alpha 16 \
|
|
--output_dir output/zen-nano-thinking-v1.0.1 \
|
|
--per_device_train_batch_size 2 \
|
|
--gradient_accumulation_steps 4 \
|
|
--learning_rate 1e-5 \
|
|
--num_train_epochs 1 \
|
|
--max_steps 100 \
|
|
--logging_steps 10 \
|
|
--save_steps 50 \
|
|
--do_train
|
|
|
|
# Export merged model
|
|
python src/export.py \
|
|
--model_name_or_path zenlm/zen-nano-instruct \
|
|
--adapter_name_or_path output/zen-nano-thinking-v1.0.1 \
|
|
--export_dir models/zen-nano-thinking-v1.0.1
|
|
|
|
echo "✅ Training complete for zen-nano-thinking-v1.0.1"
|