vdpappu/lora_scienceqa

Model Card for Model ID

A Gemma-2b finetuned LoRA trained on science Q&A

• Developed by: Venkat

How to Get Started with the Model

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig
from peft import PeftModel
from typing import Optional
import time
import os

def generate_prompt(input_text: str, instruction: Optional[str] = None) -> str:
    text = f"### Question: {input_text}\n\n### Answer: "
    if instruction:
        text = f"### Instruction: {instruction}\n\n{text}"
    return text

huggingface_token = os.environ.get('HUGGINGFACE_TOKEN')

base_model = AutoModelForCausalLM.from_pretrained("google/gemma-2b", token=huggingface_token)
tokenizer = AutoTokenizer.from_pretrained("google/gemma-2b", token=huggingface_token)

lora_model = PeftModel.from_pretrained(base_model, "vdpappu/lora_scienceqa")
merged_model = lora_model.merge_and_unload()

eos_token = '<eos>'
eos_token_id = tokenizer.encode(eos_token, add_special_tokens=False)[-1]

generation_config = GenerationConfig(
       eos_token_id=tokenizer.eos_token_id,
       min_length=5,
       max_length=200,
       do_sample=True,
       temperature=0.7,
       top_p=0.9,
       top_k=50,
       repetition_penalty=1.5,
       no_repeat_ngram_size=3,
       early_stopping=True
   )

question = "Which is the smoothest? Choose from: concrete sidewalk, sandpaper, paper."
prompt = generate_prompt(input_text=question)

with torch.no_grad():
    inputs = tokenizer(prompt, return_tensors="pt")
    output = merged_model.generate(**inputs, generation_config=generation_config)
    response = tokenizer.decode(output[0], skip_special_tokens=True)

print(response)

• PEFT 0.12.0