aapot
commited on
Commit
•
6db6916
1
Parent(s):
e415206
Update train script
Browse files- pretrain_llama_3b.sh +5 -3
pretrain_llama_3b.sh
CHANGED
@@ -24,8 +24,8 @@ python3 -m EasyLM.models.llama.llama_train \
|
|
24 |
--optimizer.type='lion' \
|
25 |
--optimizer.lion_optimizer.weight_decay=1.0 \
|
26 |
--optimizer.lion_optimizer.lr_schedule_type='warmup_constant' \
|
27 |
-
--optimizer.lion_optimizer.lr=
|
28 |
-
--optimizer.lion_optimizer.end_lr=
|
29 |
--optimizer.lion_optimizer.lr_warmup_steps=60000 \
|
30 |
--optimizer.lion_optimizer.lr_decay_steps=100000 \
|
31 |
--optimizer.lion_optimizer.bf16_momentum=True \
|
@@ -39,6 +39,8 @@ python3 -m EasyLM.models.llama.llama_train \
|
|
39 |
--train_dataset.huggingface_dataset.batch_size=64 \
|
40 |
--eval_dataset.type='huggingface' \
|
41 |
--eval_dataset.text_processor.fields='text' \
|
|
|
|
|
42 |
--eval_dataset.huggingface_dataset.path='/researchdisk/lm_training_dataset_first_stage' \
|
43 |
--eval_dataset.huggingface_dataset.split='validation' \
|
44 |
--eval_dataset.huggingface_dataset.seq_length=2048 \
|
@@ -46,7 +48,7 @@ python3 -m EasyLM.models.llama.llama_train \
|
|
46 |
--checkpointer.save_optimizer_state=True \
|
47 |
--logger.online=True \
|
48 |
--logger.prefix='EasyLM' \
|
49 |
-
--logger.project="llama-3b-
|
50 |
--logger.output_dir="gs://finnish-nlp-research-us/llama-3b-v2-checkpoint" \
|
51 |
--logger.wandb_dir="./"
|
52 |
|
|
|
24 |
--optimizer.type='lion' \
|
25 |
--optimizer.lion_optimizer.weight_decay=1.0 \
|
26 |
--optimizer.lion_optimizer.lr_schedule_type='warmup_constant' \
|
27 |
+
--optimizer.lion_optimizer.lr=1e-4 \
|
28 |
+
--optimizer.lion_optimizer.end_lr=1e-5 \
|
29 |
--optimizer.lion_optimizer.lr_warmup_steps=60000 \
|
30 |
--optimizer.lion_optimizer.lr_decay_steps=100000 \
|
31 |
--optimizer.lion_optimizer.bf16_momentum=True \
|
|
|
39 |
--train_dataset.huggingface_dataset.batch_size=64 \
|
40 |
--eval_dataset.type='huggingface' \
|
41 |
--eval_dataset.text_processor.fields='text' \
|
42 |
+
--eval_dataset.text_processor.add_eos_token=True \
|
43 |
+
--eval_dataset.text_processor.add_bos_token=True \
|
44 |
--eval_dataset.huggingface_dataset.path='/researchdisk/lm_training_dataset_first_stage' \
|
45 |
--eval_dataset.huggingface_dataset.split='validation' \
|
46 |
--eval_dataset.huggingface_dataset.seq_length=2048 \
|
|
|
48 |
--checkpointer.save_optimizer_state=True \
|
49 |
--logger.online=True \
|
50 |
--logger.prefix='EasyLM' \
|
51 |
+
--logger.project="llama-3b-v2" \
|
52 |
--logger.output_dir="gs://finnish-nlp-research-us/llama-3b-v2-checkpoint" \
|
53 |
--logger.wandb_dir="./"
|
54 |
|