sihanxu/slimpajama-qa-pool

SlimPajama question pool

• qa_pool.npy: uint16 array of shape [96488, 2048] holding Mistral-7B-v0.1 token ids. md5: 64ef7305ced71cdfcd4e679b1edee81a.

Each row is one sequence of 2048 tokens in which a document is read twice, with four questions between the two reads and the answers after the second:

[d] S Q1 S Q2 S Q3 S Q4 S [d] S Q1 A1 S Q2 A2 S Q3 A3 S Q4 A4

Here d is the start of a document (as long as the sequence allows), Q a question, A its answer, and S the separator "\n\n". A newline rounds the sequence to 2048 tokens. The rows are shuffled with a fixed seed.

How it was made

• Documents. Train shard 00009 of SlimPajama-627B, using documents of at least 1,000 Mistral tokens. The first 850 tokens of each document are shown to the question model.
• Questions. Qwen3.5-9B (greedy decoding, vLLM) wrote four questions about specific facts in each document. Each answer is a span of 1 to 6 words copied from the text. A pair is kept only if its answer occurs verbatim in the text and not in the question. A document is kept if it has four such pairs: 96,488 of about 112,000.

Use

import numpy as np
pool = np.load("qa_pool.npy", mmap_mode="r")   # [96488, 2048], Mistral-7B-v0.1 token ids

License

The document text comes from SlimPajama-627B and is subject to the licenses of its sources. The questions and answers were generated by Qwen3.5-9B.