T5 Small Tapaco

🚀 T5-small 用於釋義生成

本項目是基於 Google 的 T5-small 模型，在 TaPaCo 數據集上進行微調，以實現釋義生成功能。該模型能夠將輸入的句子轉換為多種不同表述但意思相近的句子，為文本處理提供了更多的靈活性。

🚀 快速開始

模型使用示例

以下是使用該模型進行釋義生成的代碼示例：

from transformers import T5ForConditionalGeneration, T5Tokenizer

tokenizer = T5Tokenizer.from_pretrained("hetpandya/t5-small-tapaco")
model = T5ForConditionalGeneration.from_pretrained("hetpandya/t5-small-tapaco")

def get_paraphrases(sentence, prefix="paraphrase: ", n_predictions=5, top_k=120, max_length=256,device="cpu"):
        text = prefix + sentence + " </s>"
        encoding = tokenizer.encode_plus(
            text, pad_to_max_length=True, return_tensors="pt"
        )
        input_ids, attention_masks = encoding["input_ids"].to(device), encoding[
            "attention_mask"
        ].to(device)

        model_output = model.generate(
            input_ids=input_ids,
            attention_mask=attention_masks,
            do_sample=True,
            max_length=max_length,
            top_k=top_k,
            top_p=0.98,
            early_stopping=True,
            num_return_sequences=n_predictions,
        )

        outputs = []
        for output in model_output:
            generated_sent = tokenizer.decode(
                output, skip_special_tokens=True, clean_up_tokenization_spaces=True
            )
            if (
                generated_sent.lower() != sentence.lower()
                and generated_sent not in outputs
            ):
                outputs.append(generated_sent)
        return outputs

paraphrases = get_paraphrases("The house will be cleaned by me every Saturday.")

for sent in paraphrases:
  print(sent)

輸出示例

運行上述代碼後，可能會得到以下輸出：

The house is cleaned every Saturday by me.
The house will be cleaned on Saturday.
I will clean the house every Saturday.
I get the house cleaned every Saturday.
I will clean this house every Saturday.

📚 詳細文檔

模型微調

若你想了解如何對該模型進行微調，請參考以下指南： https://towardsdatascience.com/training-t5-for-paraphrase-generation-ab3b5be151a2

作者信息

本項目由 Het Pandya/@hetpandya 創建，你還可以通過 LinkedIn 聯繫作者。

本項目在印度用心打造 ❤️