gentune-flan 0.1.8

Genetic Algorithm: Optimize the finetuning of FLAN-T5 models

gentune-flan

Genetic Algorithm used to optimise the finetunning of FLAN-T5

library dependecies

The following libraries must be installed for the library gentune-flan to work

pip install evaluate, transformers, rouge-score, pandas, numpy, torch

the gentune-flan library

pip install gentune-flan

Example (Summery creation from FLAN-T5 Base Model)

import gentune-flan

dataset = dataset = load_dataset("cnn_dailymail", "3.0.0", cache_dir="./hf_cache")

test_data = dataset["test"].select(range(100))

prompts = []

references = []

for ex in (list(test_data))[0:10]:

article = ex.get("article", "")

summary = ex.get("highlights", "")

if isinstance(article, str) and isinstance(summary, str) and article.strip() != "" and summary.strip() != "":

    prompt = f"Summarize the following article:\n\n{article.strip()}\n\nSummary:"
    
    prompts.append(prompt)
    
    references.append(summary.strip())

params = {

        "max_length": [200, 300,400,500,600,700],
        
        "temperature": [0.1,0.3,0.5,0.7,0.9,1],
        
        "top_k": [10,20,30,50,70,90,100],
        
        "top_p": [0.2,0.3,0.5,0.7,0.9],
        
        "do_sample": True,       # Enables sampling-based decoding
        
        "num_beams": 1           # Beam search disabled (greedy/sample)
        
    }

args = gentune-flan.optimize_flan_t5_base(prompts=prompts, references=references, params=params, error_metric='ROUGE-L', number_of_generation = 2, mutation_rate=0.05, population_size=4, random_seed=2025)

print(args)

model_name = "google/flan-t5-base" # or your fine-tuned path

tokenizer = AutoTokenizer.from_pretrained(model_name)

model = AutoModelForSeq2SeqLM.from_pretrained(model_name)

inputs = tokenizer(

              prompts,
              
              return_tensors="pt",
              
              padding=True,
              
              truncation=True,
              
              max_length=512
          )

with torch.no_grad():

        outputs = model.generate(
        
            input_ids=inputs["input_ids"],
            
            attention_mask=inputs["attention_mask"],
            
            **args
        )

decoded_preds = tokenizer.batch_decode(outputs, skip_special_tokens=True)

rouge = evaluate.load("rouge")

result = rouge.compute(predictions=decoded_preds, references=references, use_stemmer=True)

rouge_value = round(result['rougeL'] * 100, 2)

print(rouge_value)

Default Hyperparameter value range

max_length : [50,100,150,200,250,300]

temperature : [0.2,0.4,0.6,0.8]

top_k : [10,20,30,40,50]

top_p : [0.3,0.5,0.7,0.9]

error_metric = 'ROUGE-L' (user can check for 'ROUGE-1', 'ROUGE-2' also)

number_of_generation = 3 (user can increase the value for better evolution)

mutation_rate = 0.05

population_size = 10 (user can increase the value for larger set for better results)