TinyStories-3M-val-Hebrew
This model is trained upon Norod78/TinyStoriesV2-GPT4-valid_heb-lineByLine-EoT
Dataset is a machine translation of TinyStoriesV2-GPT4-valid.txt by roneneldan
Trasnlation was done using this script
Original Dataset containing synthetically generated (by GPT-3.5 and GPT-4) short stories that only use a small vocabulary.
Model description
A very very small model (8M params) tarined on a very small dataset
A sample inference script is available
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 0.0004
- train_batch_size: 24
- eval_batch_size: 8
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: cosine_with_restarts
- lr_scheduler_warmup_steps: 500
- num_epochs: 300.0
Framework versions
-
Transformers 4.31.0.dev0
-
Pytorch 2.0.0
-
Datasets 2.13.1
-
Tokenizers 0.13.3
-
Parameter calculation
def gpt_params(seq_len, vocab_size, d_model, num_heads, num_layers):
""" Given GPT config calculate total number of parameters """
ffw_size = 4*d_model # in GPT the number of intermediate features is always 4*d_model
# token and position embeddings
embeddings = d_model * vocab_size + d_model * seq_len
# transformer blocks
attention = 3*d_model**2 + 3*d_model # weights and biases
attproj = d_model**2 + d_model
ffw = d_model*(ffw_size) + ffw_size
ffwproj = ffw_size*d_model + d_model
layernorms = 2*2*d_model
# dense
ln_f = 2*d_model
dense = d_model*vocab_size # note: no bias here
# note: embeddings are not included in the param count!
total_params = num_layers*(attention + attproj + ffw + ffwproj + layernorms) + ln_f + dense
return total_params
#gpt2 = dict(seq_len = 1024, vocab_size = 50257, d_model = 768, num_heads = 12, num_layers = 12)
gpt2 = dict(seq_len = 256, vocab_size = 50259, d_model = 128, num_heads = 16, num_layers = 8)
result = gpt_params(**gpt2)/1e6
print(result) #Prints 8.019584