fine tune llama 3.2
- train command:
nohup python train_full.py --data_file data/sarcasm.csv \
--output_dir logs \
--num_train_epochs 3 \
--batch_size 4 \
--learning_rate 1e-6 \
--max_length 128 > output.log 2>&1 &
- train lora:
nohup python train_lora.py --data_file data/sarcasm.csv \
--output_dir logs \
--num_train_epochs 2 \
--batch_size 2 \
--learning_rate 1e-5 \
--max_length 128 \
--lora_r 16 \
--lora_alpha 32 \
--lora_dropout 0.1 > output.log 2>&1 &
-
view log:
tail -f output.log -
view
tensorboard:tensorboard --logdir logs/logs -
test model:
python test.py --data_file_path ./data/sarcasm.csv --model_id fine-tuned-model --cuda_device 0 --max_length 128 -
inference model:
python inference.py logs/checkpoint-190
before finetune
{'BERTScore': {'Precision': 0.8277986645698547, 'Recall': 0.8530203700065613, 'F1': 0.8401092886924744}, 'ROUGE-L': {'F1': 0.0794529113715608, 'Precision': 0.057397086892784416, 'Recall': 0.16667777018694505}, 'BLEU-4': 0.0027716679935327476, 'F1-Score': 0.8402202725410461}
after finetune
{'BERTScore': {'Precision': 0.8509647250175476, 'Recall': 0.876897931098938, 'F1': 0.863616406917572}, 'ROUGE-L': {'F1': 0.1416847936719217, 'Precision': 0.10668925006579812, 'Recall': 0.23828849169350153}, 'BLEU-4': 0.22531999050937368, 'F1-Score': 0.8637367486953735}
Note: RuntimeError: TensorBoardCallback requires tensorboard to be installed. Either update your PyTorch version or install tensorboardX.