Richard-Sieg-TH-Koln/anlp-tinystories-instruct-gpt2

TinyStoriesInstruct, tokenized for supervised fine-tuning

Prompt/response pairs built from roneneldan/TinyStoriesInstruct, tokenized with tiktoken's GPT-2 encoding, for the Advanced NLP block course at TH Koeln, block 07 (supervised fine-tuning).

Prompt format

Plain text, no chat template. Whichever of Summary, Features, Words, Random sentence are present in the source example, always in that order, each on its own labeled line, followed by Story:. The story is the response. Source field order is not fixed and was not preserved; a consistent prompt structure was judged more useful than faithfulness to the source's inconsistent ordering.

Files

train_tokens.npy, train_loss_mask.npy, train_attn_mask.npy, and the same three for valid. Every array is (n_examples, 512).

• tokens: uint16, GPT-2 token ids, right-padded
• loss_mask: uint8, 1 on response tokens (including the closing <|endoftext|>), 0 on prompt tokens and on padding
• attn_mask: uint8, 1 on every real token (prompt and response), 0 on padding

GPT-2 has no pad token. <|endoftext|> (id 50256) is reused as the pad value; this is safe only because attn_mask and loss_mask mark which positions are real, so the model is never trained or evaluated on a padding position pretending to be a real <|endoftext|>.

Usage

import numpy as np
from huggingface_hub import hf_hub_download

path = hf_hub_download("Richard-Sieg-TH-Koln/anlp-tinystories-instruct-gpt2", "train_tokens.npy", repo_type="dataset")
train_tokens = np.load(path)  # (n_examples, 512)