FLAN-OPT-2.7b-LoRA

FLAN-OPT-2.7b-LoRA

OPT was first introduced in Open Pre-trained Transformer Language Models and first released in metaseq's repository on May 3rd 2022 by Meta AI.

This model is facebook/opt-2.7b finetuned with low-rank adapters (https://arxiv.org/abs/2106.09685) on the FLAN datasets (https://arxiv.org/pdf/2210.11416.pdf).

Low-rank adapters (r=16) finetuned over 1.1m new tokens of a FLAN task mixture, with the start of each example cut off if it was too large to fit within a 256 token context.

The model reaches a train ppl of 5.09 and an eval ppl of 4.36.

Inference Example (Chain-of-Thought prompt):

# %pip install -qq transformers git+https://github.com/huggingface/peft accelerate bitsandbytes
from peft import PeftModel, PeftConfig
from transformers import AutoModelForCausalLM, AutoTokenizer
peft_model_id = "crumb/FLAN-OPT-2.7b-LoRA"

config = PeftConfig.from_pretrained(peft_model_id)
model = AutoModelForCausalLM.from_pretrained(config.base_model_name_or_path, load_in_8bit=True, low_cpu_mem_usage=True, device_map='auto')
model = PeftModel.from_pretrained(model, peft_model_id)
tokenizer = AutoTokenizer.from_pretrained(config.base_model_name_or_path)

import torch
prompt = """
Q: Answer the following yes/no question by reasoning step-by-step. Could a dandelion suffer from hepatitis?
A: Hepatitis only affects organisms with livers. Dandelions don’t have a liver. The answer is no.

Q: Answer the following yes/no question by reasoning step-by-step. Can you write a whole Haiku in a single tweet?
A: A haiku is a japanese three-line poem. That is short enough to fit in 280 characters. The answer is yes.

Q: Answer the following yes/no question by reasoning step-by-step. Can you reach space with a Cessna?
A: 
""".strip()
inputs = tokenizer([prompt], return_tensors='pt')

with torch.autocast("cuda", dtype=torch.float16):
    outputs = model.generate(
        input_ids=inputs.input_ids.cuda(),
        attention_mask=inputs.attention_mask.cuda(),
        max_new_tokens=32,
        top_k=4,
        penalty_alpha=0.6
    )
print("\n".join(tokenizer.decode(outputs[0]).split("\n")[:prompt.count("\n")+1]))
# Cessna is a single-engine aircraft. Cessna cannot reach space. The answer is no.