Details
Hi, I've long been looking for a high-quality Wikipedia embeddings dataset made with a SOTA model, so your repo is of great interest to me.
Could you explain in some more detail how the embeddings were generated? Most importantly, what text exactly was processed for each article, the raw Wikitext or some parsed version? Were links/categories/references stripped? I'm asking because the "snippet" column appears to contain garbage for some articles (especially redirects and disambiguations), which is already visible in the preview.
Also, Qwen3-Embedding-4B has a maximum sequence length of 32k tokens, but many Wikipedia articles are substantially longer than that. How did you deal with this?
And how long did this take to generate? I've generated embeddings for just the first paragraph of all English Wikipedia pages using much smaller models on a few occasions, and I saw multi-hour runtimes (on a modest GPU). Did you process this with a cluster?