# "-2" is hardcoded for the Llama tokenizer to make the offset correct.?

Nobody has claimed this yet.

In the main training file, I saw that you applied a -2 to the instruction length and the comment is that -2 is hardcoded for the right offset. Can you help me with the exact reason to do so?

I cross-checked with the llama-recipe library
https://github.com/facebookresearch/llama-recipes/blob/main/src/llama_recipes/datasets/alpaca_dataset.py

In the following codeblock, no offset was applied.

` def __getitem__(self, index):
IGNORE_INDEX = -100 # The default setting in CrossEntropyLoss

ann = self.ann[index]
if ann.get("input", "") == "":
prompt = PROMPT_DICT["prompt_no_input"].format_map(ann)
else:
prompt = PROMPT_DICT["prompt_input"].format_map(ann)
example = prompt + ann["output"]
prompt = torch.tensor(
self.tokenizer.encode(prompt), dtype=torch.int64
)
example = self.tokenizer.encode(example)
example.append(self.tokenizer.eos_token_id)
example = torch.tensor(
example, dtype=torch.int64
)
labels = copy.deepcopy(example)
labels[: len(prompt)] = -1
example_mask = example.ge(0)
label_mask = labels.ge(0)
example[~example_mask] = 0
labels[~label_mask] = IGNORE_INDEX

return {
"input_ids": example.tolist(),
"labels": labels.tolist(),
"attention_mas`

Start in the main training file at the instruction-length calculation containing the hardcoded -2, then compare its tokenizer behavior with the linked llama-recipes alpaca_dataset.py implementation. Establish why the offset is needed and document that rationale, or record the evidence for removing it; the issue is complete when the offset behavior and its intended effect are clearly explained.

Written by the indexing model from the issue text.