File size: 2,297 Bytes
6841785 94ad50a 0158d31 94ad50a 0158d31 94ad50a |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 |
---
license: apache-2.0
---
This is a demo of how to pretrain a mistral architecture model by SFT Trainer ,and it needs only 70 lines Python code.
```
import torch
from transformers import TrainingArguments, MistralForCausalLM, MistralModel, MistralConfig, AutoTokenizer
from datasets import load_dataset
from trl import SFTTrainer
configuration = MistralConfig(vocab_size=32000,
hidden_size=2048,
intermediate_size=7168,
num_hidden_layers=24,
num_attention_heads=32,
num_key_value_heads=8,
hidden_act="silu",
max_position_embeddings=4096,
pad_token_id=2,
bos_token_id=1,
eos_token_id=2)
model = MistralForCausalLM(configuration)
#model = MistralForCausalLM.from_pretrained("./6B_code_outputs/checkpoint-10000")
tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-Instruct-v0.2", local_files_only=False)
tokenizer.pad_token = tokenizer.eos_token
dataset = load_dataset('HuggingFaceTB/cosmopedia-20k', split="train")
#dataset = load_dataset('Elriggs/openwebtext-100k', split="train")
dataset = dataset.shuffle(seed=42)
print(f'Number of prompts: {len(dataset)}')
print(f'Column names are: {dataset.column_names}')
def create_prompt_formats(sample):
"""
Format various fields of the sample ('instruction', 'context', 'response')
Then concatenate them using two newline characters
:param sample: Sample dictionnary
"""
output_texts = []
for i in range(len(sample['text'])):
formatted_prompt = sample['text'][i]
output_texts.append(formatted_prompt)
#print(output_texts)
return output_texts
trainer = SFTTrainer(
model,
train_dataset=dataset,
tokenizer = tokenizer,
max_seq_length=2048,
formatting_func=create_prompt_formats,
args=TrainingArguments(
per_device_train_batch_size=2,
gradient_accumulation_steps=1,
warmup_steps=2,
max_steps=10000,
learning_rate=1e-4,
logging_steps=1,
output_dir="1B_outputs", overwrite_output_dir=True,save_steps=1000,
optim="paged_adamw_32bit",report_to="none"
)
)
trainer.train()
trainer.model.save_pretrained("1B-final", dtype=torch.float32)
trainer.tokenizer.save_pretrained("1B-final")
``` |