zitaos/online-source-inversion-data

Online source inversion: data bundle

Everything the online source inversion server needs at run time, except the model weights: the search indexes of ~1M pretraining documents, the BM25 index, and the documents' texts. Code and instructions: the online-source-inversion repository (README, "Install the server side").

Use

Download the whole bundle (39 GB) and point the server at its u1m/ folder:

hf download zitaos/online-source-inversion-data --repo-type dataset --local-dir si-online-bundle
# then, in the repository's config/server.env
export SI_ROOT=/path/to/si-online-bundle/u1m
export HF_HOME=/path/to/hf_home      # the model weights: allenai/OLMo-2-1124-7B (downloaded from the Hugging Face Hub)

Check the download with sha256sum -c SHA256SUMS (in si-online-bundle/). On the first start the server writes a BM25 cache (u1m/data/u/bm25/cache/, ~3 GB) and its logs (u1m/results/scaleup/online/) into the folder, so it must be writable.

Contents

Path under u1m/	Size	What
results/scaleup/online/index/olmo2_7b_final__32pre__index/	8.2 GB	OLMo-2 7B URL-centroid index: one vector per source URL (998,647), from the 7B's hidden states at layer 32pre (z-scored)
results/scaleup/online/index/qwen3emb8b__final__index/	8.2 GB	Qwen3-Embedding-8B URL-centroid index (whitened), for the optional qwen3 method
data/u/bm25/	1.6 GB	Document-level BM25 index (k1 = 3.0, b = 1.0) and its vocabulary; keep the three files together
data/docs.parquet	7.5 GB	1.15M documents: id, URL, origin, full text
data/chunks.parquet, data/splits.parquet	6.7 GB	5.9M chunks of ≤ 2,000 characters (the first ≤ 8 of each document) and their roles
data/u/doctext/	5.7 GB	The documents' full texts exported for random access (zstd per document)
results/scaleup/layers/*/summary.json, results/scaleup/lexical/bm25_doc_tuning.json	small	The post-processing chosen per layer, and the BM25 parameters

Each index's meta.json records what it was built from. An index is valid only for the model checkpoint and layer it was built from; document ids mean the same document everywhere in the bundle.

Where the data comes from

The documents are a uniform random sample of OLMo-Mix-1124 (allenai/olmo-mix-1124), the pretraining data of OLMo-2: DCLM web text, peS2o, arXiv, Wikipedia, StarCoder, OpenWebMath and Algebraic Stack. 1M documents form the index; 100k more are held out. The indexes were built from the hidden states and embeddings of the 1M documents' chunks.

Terms

The document texts come from OLMo-Mix-1124, released under the Open Data Commons Attribution License (ODC-By) v1.0, and remain subject to the terms of their original sources (e.g. Common Crawl for the web text). Use the texts accordingly, and cite OLMo-Mix-1124 / OLMo 2 when you use them.