llama

[yangYJT]

can this project support llama?

[ningding97]

Of course, we have included a tutorial to use GPT-J to train Chat AI. You can use similar approach to train Llama, for example

from openprompt import plms
from openprompt.plms import *
from transformers import GPTJConfig, GPTJModel, GPTJForCausalLM, GPT2Tokenizer, LlamaConfig, LlamaForCausalLM, LlamaTokenizer
plms._MODEL_CLASSES["gptj"]= ModelClass(**{"config": GPTJConfig, "tokenizer": GPT2Tokenizer, "model": GPTJForCausalLM, "wrapper": LMTokenizerWrapper})
plms._MODEL_CLASSES["llama"]= ModelClass(**{"config": LlamaConfig, "tokenizer": LlamaTokenizer, "model": LlamaForCausalLM, "wrapper": LMTokenizerWrapper})

Please note that we are not distributing Llama checkpoints.

[shuaizhao95]

According to your method in a classification task, the output logits are equal. Could you help me take a look at it?

[shuaizhao95]

How to solve the logits equality of LLaMA output?

This is my code.

from datasets import load_dataset
from transformers import set_seed
from openprompt.data_utils import InputExample
import os
from tqdm import tqdm

device = "cuda"
classes = ["negative", "positive"]
set_seed(1024)
from accelerate import Accelerator

accelerator = Accelerator()

data_path = 'data'
test_path = os.path.join(data_path, 'test.json')
test_dataset = load_dataset('json', data_files=test_path)['train'] # 1 positive 0 negative
y_true = test_dataset['label']
dataset = []
import copy
data = []
copy_test_dataset = copy.deepcopy(test_dataset)
for example in copy_test_dataset:
temp_data = {"guid": example["label"], "text_a": example["sentence"]}
data.append(temp_data)

for item in data:
dataset.append(InputExample(guid=item["guid"], text_a=item["text_a"]))

from openprompt import plms
from openprompt.plms import *
from transformers import LlamaConfig, LlamaForCausalLM, LlamaTokenizer
plms._MODEL_CLASSES["llama"]= ModelClass(**{"config": LlamaConfig, "tokenizer": LlamaTokenizer, "model": LlamaForCausalLM, "wrapper": LMTokenizerWrapper})

from openprompt.plms import load_plm
plm, tokenizer, model_config, WrapperClass = load_plm("llama", "huggyllama/llama-7b")
tokenizer.pad_token_id = 0

from openprompt.prompts import ManualTemplate

promptTemplate = ManualTemplate(
text=' {"placeholder":"text_a"} This sentence was {"mask"}',
tokenizer=tokenizer,
)

from openprompt.prompts import ManualVerbalizer
promptVerbalizer = ManualVerbalizer(classes=classes,
label_words={"negative": ["bad"], "positive": ["good", "wonderful", "great"], },
tokenizer=tokenizer, )

from openprompt import PromptForClassification
promptModel = PromptForClassification(template=promptTemplate, plm=plm, verbalizer=promptVerbalizer, )
from openprompt import PromptDataLoader

data_loader = PromptDataLoader(dataset=dataset, tokenizer=tokenizer, template=promptTemplate,
tokenizer_wrapper_class=WrapperClass, batch_size=1)

import torch

promptModel.eval()
print(promptModel)
promptModel, data_loader = accelerator.prepare(promptModel, data_loader)
promptModel.to(device)
predictions = []
with torch.no_grad():
for batch in tqdm(data_loader, desc="Processing batches"):
batch = {k: v.to(device) for k, v in batch.items()}
print(batch)
logits = promptModel(batch)
print(logits)
exit()
preds = torch.argmax(logits, dim=-1)
for i in preds:
predictions.append(i.item())

from sklearn.metrics import accuracy_score

accuracy = accuracy_score(y_true, predictions)
print('Accuracy: %.2f' % (accuracy * 100))

The output logits is :

tensor([[-1.3863, -1.3863]])