Special tokens not loaded / data_collator not modified

[banank1989]

In your finetuning code at:

https://colab.research.google.com/drive/1n5U13L0Bzhs32QO_bls5jwuZR62GPSwE?usp=sharing

As per dolly code, special tokens needs to be loaded to tokenizer

    tokenizer.pad_token = tokenizer.eos_token
    tokenizer.add_special_tokens(
        {"additional_special_tokens": [END_KEY, INSTRUCTION_KEY, RESPONSE_KEY_NL]}
    )

Also, you have not modified data_collator

 DataCollatorForCompletionOnlyLM(DataCollatorForLanguageModeling):
    def torch_call(self, examples):
        batch = super().torch_call(examples)

        # The prompt ends with the response key plus a newline
        response_token_ids = self.tokenizer.encode(RESPONSE_KEY_NL)
        labels = batch["labels"].clone()

        for i in range(len(examples)):
            response_token_ids_start_idx = None
            for idx in np.where(batch["labels"][i] == response_token_ids[0])[0]:
                response_token_ids_start_idx = idx
                break

            if response_token_ids_start_idx is None:
                raise RuntimeError(
                    f'Could not find response key {response_token_ids} in token IDs {batch["labels"][i]}'
                )

            response_token_ids_end_idx = response_token_ids_start_idx + 1

            # loss function ignore all tokens up through the end of the response key
            labels[i, :response_token_ids_end_idx] = -100

        batch["labels"] = labels

        return batch

data_collator = DataCollatorForCompletionOnlyLM(
        tokenizer=tokenizer, mlm=False, return_tensors="pt", pad_to_multiple_of=8
    )

I am assumig they are also needed? Or Am I missing something?