Dataset Viewer

The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.

zotero-rag gold evaluation corpus

Source documents, a prebuilt vector index and published results for the rag-quality-eval skill of zotero-rag. The corpus is the public Zotero group test-rag-plugin (about reference management software).

Contents

Path What
sources/ the group's open-access PDFs (<attachmentKey>.pdf) and manifest.json (item metadata)
snapshots/<embedding-model>/ latest.jsonl.gz: chunks, vectors and index metadata exported with bin/library_snapshot.py; latest.json: how it was built (Kreuzberg extractor; another extractor's snapshot is in a subfolder named after it, e.g. docling/)
reports/ one folder per published eval run, plus index.json
evaluations/ one-off studies (data, scripts and a README with the findings), one folder each

License and attribution

The PDFs are open-access publications licensed CC BY 4.0; copyright remains with their authors. sources/manifest.json lists title, authors, year and the Zotero item of each. The snapshot contains the documents' text as chunks and is therefore distributed under the same license. Gold questions live in the code repository (skills/rag-quality-eval/gold/).

Use

uv run python skills/rag-quality-eval/scripts/gold_env.py up restores the index from cmboulanger/zotero-rag-gold into a local backend; see skills/rag-quality-eval/SKILL.md.

Sources

39 PDFs, fingerprint 7e0996940ed8.

Index snapshots

Embedding model Chunks Extractor Preset Code commit Built
multilingual-e5-large-instruct (1024-dim) 12019 kreuzberg remote-kisski 2f7b574 2026-10-12

Evaluations

Study Folder
Extractor comparison: Docling vs Kreuzberg (2026-10-14) evaluations/extractor-comparison-20261014/

Results

Latest run: 20261013T054355Z (report); every run is kept under reports/<run>/.

Composite is the mean over runs (0-1); recall is fact recall; judge is the 0-5 judged overall.

Run Gold Commit Preset / model Runs Composite Recall en Recall non-en Judge Lost retr. Lost gen.
20261014-extractor-cmp-kfresh v3 42f7ccc remote-kisski / gemma-4-31b-it 33 0.84 0.69 0.44 3.24 34 8
20261014-extractor-cmp-kfresh v3 42f7ccc remote-kisski / qwen3-30b-a3b-instruct-2507 33 0.81 0.62 0.46 3.18 34 7
20261014-extractor-cmp-docling v3 42f7ccc kisski-docling / gemma-4-31b-it 33 0.81 0.43 0.78 2.97 38 5
20261014-extractor-cmp-docling v3 42f7ccc kisski-docling / qwen3-30b-a3b-instruct-2507 33 0.80 0.50 0.69 3.06 35 5
20261014-chunk1600-kreuzberg v3 aa3f320 rk-c1600 / gemma-4-31b-it 33 0.85 0.63 0.66 3.39 18 15
20261014-chunk1600-kreuzberg v3 aa3f320 rk-c1600 / qwen3-30b-a3b-instruct-2507 33 0.79 0.53 0.61 3.36 17 24
20261014-chunk1600-docling v3 aa3f320 kd-c1600 / gemma-4-31b-it 33 0.82 0.60 0.59 3.18 32 7
20261014-chunk1600-docling v3 aa3f320 kd-c1600 / qwen3-30b-a3b-instruct-2507 33 0.78 0.60 0.48 2.94 27 20
20261013T054355Z v3 ea45593 remote-kisski / gemma-4-31b-it 16 0.84 0.51 0.57 2.93 19 0
20261013T054355Z v3 ea45593 remote-kisski / qwen3-30b-a3b-instruct-2507 16 0.78 0.49 0.33 2.57 19 0
20261012T155733Z v2 969f1b8 remote-kisski / gemma-4-31b-it 10 0.79 0.58 0.52 2.90 11 0
20261012T155733Z v2 969f1b8 remote-kisski / qwen3-30b-a3b-instruct-2507 10 0.73 0.47 0.49 2.60 11 5
20261012T155312Z v2 969f1b8 remote-kisski / gemma-4-31b-it 10 0.77 0.37 0.52 2.30 15 0
20261012T155312Z v2 969f1b8 remote-kisski / qwen3-30b-a3b-instruct-2507 10 0.67 0.35 0.44 2.30 14 2
20261012T131708Z v2 969f1b8 remote-kisski / gemma-4-31b-it 10 0.73 0.28 0.46 1.90 17 0
20261012T131708Z v2 969f1b8 remote-kisski / qwen3-30b-a3b-instruct-2507 10 0.67 0.26 0.44 2.10 19 1
Downloads last month
564