Hugging Face's logo

Spaces:
nroggendorff
/
train-llama
Runtime error

App Files Community
34
nroggendorff commited on
Commit
d82d31b
·
verified ·
1 Parent(s): d5968da

Update train.py

Browse files
Files changed (1) hide show
  1. train.py +3 -3
train.py CHANGED
@@ -8,16 +8,16 @@ from datasets import load_dataset, DatasetDict, Dataset
8
  from tokenizers import ByteLevelBPETokenizer
9
 
10
  MAX_SEQ_LENGTH = 2048
11
- BATCH_SIZE = 1
12
  EPOCHS = 2
13
  LEARNING_RATE = 5e-5
14
- FACTOR = 2048
15
  VOCAB_SIZE = 32000
16
  INPUT_DATASET = "HuggingFaceTB/smollm-corpus"
17
  INSTRUCT_DATASET = "nroggendorff/openhermes"
18
  OUTPUT_REPO = "smallama"
19
  FP16 = True
20
- WARMUP_STEPS = 200
21
  DECAY = 1e-4
22
  GRADIENT_ACCUMULATION_STEPS = 16
23
  CLIPPING = 1.0
 
8
  from tokenizers import ByteLevelBPETokenizer
9
 
10
  MAX_SEQ_LENGTH = 2048
11
+ BATCH_SIZE = 8
12
  EPOCHS = 2
13
  LEARNING_RATE = 5e-5
14
+ FACTOR = 1024
15
  VOCAB_SIZE = 32000
16
  INPUT_DATASET = "HuggingFaceTB/smollm-corpus"
17
  INSTRUCT_DATASET = "nroggendorff/openhermes"
18
  OUTPUT_REPO = "smallama"
19
  FP16 = True
20
+ WARMUP_STEPS = 20
21
  DECAY = 1e-4
22
  GRADIENT_ACCUMULATION_STEPS = 16
23
  CLIPPING = 1.0