Skip to content

Releases

QSA weights

QSA weights Pre-release
Pre-release

Choose a tag to compare

@caciolai caciolai released this 09 Nov 16:05
5ebea73

QSA indexers for Qwen3.8-27B, distilled on the frozen model - the Block Sparse Attention (Flat) baseline on Qwen3.8-27B

Agent traces

Agent traces Pre-release
Pre-release

Choose a tag to compare

@caciolai caciolai released this 02 Nov 13:33

Scrubbed agent traces for the LongDocs games and edgebench sources, as parquet shards. CC BY 4.0.

Columns: doc_id, source, license, n_tokens, text, and for edgebench also n_tokens_as_generated, model, task, run. Each source's manifest.json has the scrub rules, counts and shard checksums.

The text differs slightly from the corpus used in the paper: games leaves out 21 documents generated by Claude Opus, and edgebench has per-session harness tokens and email addresses redacted, and internal harness and tool names, paths and cluster hostnames replaced by neutral placeholders.

N-LSA weights

N-LSA weights Pre-release
Pre-release

Choose a tag to compare

@caciolai caciolai released this 02 Nov 13:13

N-LSA (and baselines) weights

LongDocs lock files v1

Pre-release

Choose a tag to compare

@caciolai caciolai released this 02 Nov 13:33

Per-document sha256 lock files for the LongDocs hub sources (arxiv, doab, gutenberg, hansard, usgpo, wikimedia, swe, code).
They contain no document text.
A lock file lists each document's id, text sha256 and position, so a collection you run yourself can be verified with --verify and a mismatch shows which documents differ.
Total 134,695 documents across the eight sources.
The paper reports 134,740 for native training; the 45 difference has not been investigated.
The code lock reflects a collection of the Stack streamed without a pinned revision, so a later collection may differ.

D-LSA weights

D-LSA weights Pre-release
Pre-release

Choose a tag to compare

@caciolai caciolai released this 02 Nov 13:13

D-LSA index maps for Qwen3.8-27B and GLM-5.3-Flash

RULER haystack sources

Pre-release

Choose a tag to compare

@caciolai caciolai released this 30 Oct 22:20

The fixed haystack sources RULER is prepared from, so lsa data prepare ruler rebuilds the paper's sets byte for byte.

File sha256 Source License
PaulGrahamEssays.json 58e352531a80cef2d22c205dbebfbfd64a8afe55a32434de845f200718756c65 Paul Graham's essays, as NVIDIA RULER's download script collected them the essays' own; included only to pin the haystack
squad.json 80a5225e94905956a6446d296ca1093975c4d3b3260f1d6c8f68bc2ab77182d8 SQuAD 2.0 dev set CC BY-SA 4.0

The essays change upstream over time, so a fresh download no longer matches the haystack the paper's sets were built from.