Update train_flan_t5.py
Browse files- train_flan_t5.py +3 -47
train_flan_t5.py
CHANGED
@@ -1,47 +1,3 @@
|
|
1 |
-
|
2 |
-
|
3 |
-
|
4 |
-
model_checkpoint = "google/flan-t5-large"
|
5 |
-
output_dir = "./finetuned-flan-t5"
|
6 |
-
|
7 |
-
# Load dataset
|
8 |
-
dataset = load_dataset("json", data_files={"train": "train_data.jsonl"})
|
9 |
-
|
10 |
-
# Tokenizer
|
11 |
-
tokenizer = AutoTokenizer.from_pretrained(model_checkpoint)
|
12 |
-
|
13 |
-
def preprocess_function(examples):
|
14 |
-
inputs = examples["input"]
|
15 |
-
targets = examples["output"]
|
16 |
-
model_inputs = tokenizer(inputs, max_length=512, truncation=True)
|
17 |
-
labels = tokenizer(targets, max_length=128, truncation=True)
|
18 |
-
model_inputs["labels"] = labels["input_ids"]
|
19 |
-
return model_inputs
|
20 |
-
|
21 |
-
tokenized_datasets = dataset.map(preprocess_function, batched=True)
|
22 |
-
|
23 |
-
# Model
|
24 |
-
model = AutoModelForSeq2SeqLM.from_pretrained(model_checkpoint)
|
25 |
-
|
26 |
-
# Training arguments
|
27 |
-
training_args = Seq2SeqTrainingArguments(
|
28 |
-
output_dir=output_dir,
|
29 |
-
evaluation_strategy="no",
|
30 |
-
learning_rate=5e-5,
|
31 |
-
per_device_train_batch_size=2,
|
32 |
-
num_train_epochs=3,
|
33 |
-
weight_decay=0.01,
|
34 |
-
save_total_limit=2,
|
35 |
-
push_to_hub=False
|
36 |
-
)
|
37 |
-
|
38 |
-
trainer = Seq2SeqTrainer(
|
39 |
-
model=model,
|
40 |
-
args=training_args,
|
41 |
-
train_dataset=tokenized_datasets["train"]
|
42 |
-
)
|
43 |
-
|
44 |
-
trainer.train()
|
45 |
-
|
46 |
-
model.save_pretrained(output_dir)
|
47 |
-
tokenizer.save_pretrained(output_dir)
|
|
|
1 |
+
tokenizer_config.json
|
2 |
+
special_tokens_map.json
|
3 |
+
spiece.model (for T5 models)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|