SLD reconstructed mini-DSTs from the 1996-1998 SLC runs and documentation corpus

This record contains two components from the SLD experiment at the SLAC Linear Collider (SLC): the reconstructed mini-Data Summary Tapes (mini-DSTs) covering the full 1996–1998 data-taking periods, and an AI-ready corpus of digitized SLD-era internal documentation. This record supersedes the earlier release of the 1996 and partial 1997 data.

The collision events are made available in two formats: the native binary “jazelle” columnar format used by the SLD reconstruction software, and a derived Apache Parquet translation that is convenient for modern columnar analysis tools (PyArrow, pandas, awkward, polars, ...). The two formats contain identical physics content. The Parquet version is the one most users will want.

Reconstructed data

68 mini-DST files covering the 1996, 1997, and 1998 runs. After translation and reading via the jazelle Python package, the dataset contains 658,241 reconstructed events. The center-of-mass energy is at the Z pole (√s ≈ 91.2 GeV) and the electron beam was longitudinally polarized, with mean magnitude 76.14% for the 1996 run and 73.08% for the 1997–1998 runs. Per-event polarization values are recorded in the PHBM bank.

Bank families

Reconstruction outputs are organized into “bank families”, each describing one detector subsystem or one stage of the reconstruction chain. The banks present in this release:

• IEVENTH — event header (run/event ID, trigger mask, timestamp).
• PHBM — beam information (beam energy, beam polarization magnitude and sign, interaction-point position and uncertainty).
• PHPSUM — inclusive list of reconstructed physics particles (4-momentum, charge, origin coordinates).
• PHCHRG — charged-track helix parameters, error matrices, hit counts, fit quality, and dE/dx from the Central Drift Chamber (CDC) and VXD3 vertex detector.
• PHKLUS — calorimeter clusters from the Liquid Argon Calorimeter (LAC), with raw and per-layer energies.
• PHCRID — Cherenkov ring likelihoods for particle identification from the Cherenkov Ring Imaging Detector (CRID).
• PHWIC — muon tracking in the Warm Iron Calorimeter (WIC).
• PHKELID — electron identification from track-calorimeter matching.
• PHPOINT — master pointer bank linking each PHPSUM entry to its associated track (PHCHRG) and/or calorimeter cluster (PHKLUS), enabling per-particle access to detector-level quantities.
• PHKCHRG — track-to-cluster matching kinematics.

The mini-DSTs are the reduced-format event records produced by the SLD reconstruction. They contain enough information for typical physics analyses (event shapes, electroweak asymmetries, jet clustering) but not the full raw detector data. Only the standard reconstruction-level data-quality requirements suppressing non-collision backgrounds are applied. No channel-specific event selection is pre-applied.

Further documentation on the jazelle file format and the field-level contents of each bank can be found in the README of the jazelle package GitHub repository (see “Related identifiers” for a link).

Sample statistics

Total events: 658,241 (102,699 from 1996, 555,542 from 1997–1998)
Hadronic Z → qq̄: 385,583 (2000 A_LR selection)
Z → e⁺e⁻: 16,547 (2001 leptonic selection)
Z → μ⁺μ⁻: 13,992 (2001 leptonic selection)
Z → τ⁺τ⁻: 17,133 (2001 leptonic selection)
Mean particles / event: 54.4
Mean charged particles / event: 13.4
Mean |P_e|: 76.14% (1996), 73.08% (1997–1998)

AI-ready documentation

A corpus of approximately 1,190 digitized documents from the SLD collaboration and the SLC accelerator program, scanned from physical copies in the SLAC archives. Over 85% date from 1980–1988, covering the R&D, design, and construction phases of the experiment: SLC machine-physics notes (beam dynamics, wakefield mitigation, damping ring lattices, klystron timing) and SLD detector hardware specifications. These notes provide context for interpreting the dataset that exists nowhere else in digital form.

Alongside the scanned page images, the release includes machine-readable text extracted by four document-processing tools (Marker, Docling, Nougat, and the Azure AI Document Intelligence API), so the corpus can serve directly as a knowledge base for retrieval-augmented and agentic LLM workflows. The extractors have complementary strengths across equations, tables, figures, and OCR quality, and their outputs are provided side by side with per-document provenance.

The corpus is organized as one directory per document, each containing:

• original.pdf — the scanned source document.
• text_marker.md, text_docling.txt, text_azure.txt, text_nougat.txt — the raw output of each extractor. The extractors available for a given document are listed in its metadata.json.
• metadata.json — document ID, source filename, available extractors, and the derived fields below.
• title.txt — the original title and an LLM-generated descriptive title.
• gist.txt — a one-sentence LLM-generated gist.
• summary.txt — a one-paragraph LLM-generated summary.
• description.txt — an LLM-generated description of the document’s visual and structural content.

Files in this record

sld_minidsts_jazelle_1996_1998.zip
68 files in the native binary jazelle format. This is a faithful preservation copy of the original SLD reconstruction output. The bank definitions necessary to interpret the format are documented in the jazelle Python reader (see “Related identifiers”). Most users should prefer the Parquet release unless they specifically need the binary format.

sld_minidsts_parquet_1996_1998.zip
68 Parquet files (one per source mini-DST) carrying the same events translated into Apache Parquet. Each file is a single Awkward record array with a top-level field per bank family.

sld_corpus.zip
The digitized documentation corpus described above: scanned SLD-era internal notes with extracted text, metadata, and LLM-generated summaries, one directory per document.

Reading the data

The Parquet files can be read directly with any Parquet-aware library:

import jazelle as jz
data = jz.from_parquet("qf....parquet")
data.PHPSUM.fields
>> ['px', 'py', 'pz', 'charge', ...]

The jazelle binary files require the jazelle Python package (see “Related identifiers” for a link). The package can also serialize the data to HDF5 or Feather. The companion sld_resurrect package wraps both formats and provides helpers for particle 4-vector construction, event-shape observables, and the published SLD event selections (A_LR, leptonic asymmetries, ...).

Provenance

The mini-DSTs were produced by the SLD reconstruction software during the 1996–1998 SLC runs. The jazelle release here is a faithful copy of those preserved files, and the Parquet release is the deterministic output of running the jazelle Python package on each one (no information is lost in the translation; field names follow the bank-internal labels). The documentation corpus was scanned from physical copies held in the SLAC archives. The extracted text is the direct output of the four named tools without manual correction, and the titles, gists, summaries, and descriptions are LLM-generated.