README.md

metadata

datasets: CarperAI/openai_summarize_tldr

from datasets import load_dataset

from transformers import AutoTokenizer, AutoModelForCausalLM

dataset = load_dataset("CarperAI/openai_summarize_tldr")

val_prompts = [sample["prompt"] for sample in dataset["valid"]]

kwargs = {
    "max_new_tokens": 50,
    "do_sample": True,
    "top_k": 0,
    "top_p": 0.95,
    "temperature": 0.5
}

model = AutoModelForCausalLM.from_pretrained("pvduy/ppo_pythia6B_sample")
model.eval()
tokenizer = AutoTokenizer.from_pretrained("pvduy/ppo_pythia6B_sample")
tokenizer.pad_token_id = tokenizer.eos_token_id

count = 0

for prompt in val_prompts:
    output_tk = tokenizer(prompt, return_tensors="pt")
    outputs = model.generate(output_tk.input_ids, attention_mask=output_tk.attention_mask, **kwargs)
    print("Prompt:", prompt)
    print("Output:", tokenizer.decode(outputs[0], skip_special_tokens=True).split("TL;DR:")[1].strip())
    print("=================================")
    count += 1
    if count == 10:
        break