PaltaTest / README.md
AlexWortega's picture
Update README.md
b4833bc
|
raw
history blame
2.08 kB
metadata
datasets:
  - Open-Orca/OpenOrca
language:
  - en
library_name: transformers

Test task

For model inference run following

from peft import PeftModel, PeftConfig
from transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig
from peft import PeftModel

seed_value = 42
torch.manual_seed(seed_value)
torch.cuda.manual_seed_all(seed_value)


model_name = "lmsys/vicuna-7b-v1.5"
lora_name = 'AlexWortega/PaltaTest'

tokenizer = LlamaTokenizer.from_pretrained(model_name, model_max_length=1024)
tokenizer.pad_token = tokenizer.eos_token

model = PeftModel.from_pretrained(
    model,
    lora_name,
    torch_dtype=torch.float16
)
model.eval()

model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float16 ).to('cpu')


model = PeftModel.from_pretrained(model, path_adapter)
model.to(device)
model.eval()

def process_output(i, o):
    """
    Simple output processing
    """
    if isinstance(o, list):
        return [seq.split('A:')[1] for seq in o]
    elif isinstance(o, str):
        return o.split('A:')[1]
    else:
        return "Unsupported data type. Please provide a list or a string."

def generate_seqs(q, k=2):
        
    q = 'Q:'+ q + 'A:'
    tokens = tokenizer.encode(q, return_tensors='pt').to(device)
    g = model.generate(input_ids=tokens)
    generated_sequences = tokenizer.batch_decode(g, skip_special_tokens=True)
    
    return  generated_sequences

q = """Given a weather description in plain text, rewrite it in a different style

```The weather is sunny and the temperature is 20 degrees. The wind is blowing at 10 km/h.
Citizens are advised to go out and enjoy the weather. The weather is expected to be sunny tomorrow.

And the following style: "Angry weatherman" """

s = generate_seqs(q=q) s = process_output(q,s) print(s[0])#


should output something like these

"""
Angry weatherman: "The weather is sunny and the temperature is 20 degrees. The wind is blowing at 10 km/h.
Citizens are advised to stay indoors and avoid going out. The weather is expected to be sunny tomorrow.

"""