marcjon/ckasketch-calibration-v1

ckasketch v1 calibration corpus — text track

1053 text documents assembled from 9 permissively-licensed sources. The fixed, hash-pinned probe set that ckasketch feeds through models to produce comparable activation-mode sketches.

• Corpus version: v1.0 (text track)
• Frozen: 2026-05-17
• Hash: sha256:cbd6a314d904842e1c5cda65eca146f8b7ddcd2027c6d0036f789a6d0a405c37 — the activation-comparability key per DESIGN.md
• Compilation license: ODC-BY 1.0
• Per-item licenses: retained from the original source — see NOTICES.md and the file list below

What this is

Activation-mode CKA (the math at the core of ckasketch) requires that any two models being compared see identical input documents. Different inputs → different activations → meaningless similarity. So we ship a fixed, reproducible probe set: this corpus.

Two sketches are comparable in activation mode iff they share (track, corpus_hash, version). — DESIGN.md §1

The corpus is input data, not the product. ckasketch's product is the per-model .sketch files (mirrored to marcjon/ckasketch-sketches). This corpus is what produced their activation arrays.

How to use

If you're generating new ckasketch activation sketches and want them to compare against existing public sketches, you MUST use this exact corpus (matching corpus_hash).

from huggingface_hub import hf_hub_download
from ckasketch.core.activation_sketch import (
    CalibrationCorpus, extract_activation_sketch,
)

corpus_path = hf_hub_download(
    repo_id="marcjon/ckasketch-calibration-v1", repo_type="dataset",
    filename="corpus.jsonl",
)
corpus = CalibrationCorpus.from_jsonl(corpus_path, track="text", version="v1")
assert corpus.corpus_hash == "cbd6a314d904842e1c5cda65eca146f8b7ddcd2027c6d0036f789a6d0a405c37", (
    "corpus_hash mismatch — sketches built from this corpus won't be "
    "comparable with public v1 sketches"
)

sketch = extract_activation_sketch(
    model_path="path/to/your/model",
    corpus=corpus,
    pooling_modes=("mean",),
    output_path="my_model.sketch",
    projection_dim=1024, projection_seed=42,
)

Source breakdown

Source	Items	Per-item license
the_stack_v2	256	BSD-3-Clause
wikipedia	175	CC-BY-SA-4.0
arxiv	128	CC-BY-4.0
openassistant	128	Apache-2.0
schema_org	128	CC-BY-SA-3.0
gutenberg	110	PD
pubmed_oa	64	CC-BY-4.0
gsm8k	32	MIT
math_dataset	32	MIT

Dataset structure

marcjon/ckasketch-calibration-v1/
├── README.md              this datacard
├── corpus.jsonl           1053 text extracts (one per line, JSON: {"id": ..., "text": ...})
├── manifest.yaml          per-item provenance + license + sha256
├── manifest.schema.yaml   JSON Schema validating every manifest entry
├── NOTICES.md             rendered per-item attribution catalog
└── CORPUS_LOCK            hash + freeze metadata

corpus.jsonl format: one JSON object per line. Each has at minimum:

• id — stable identifier within the corpus (preserves cross-model alignment)
• text — the actual text content (768-character extracts, normalized)

Loading: use ckasketch.core.activation_sketch.CalibrationCorpus.from_jsonl (see Usage section). The loader computes corpus_hash on read and verifies against this dataset's published value.

Dataset creation

Producer: the ckasketch.calibration.build pipeline.

• Fetcher modules pull from each source via official API (HuggingFace datasets, OAI-PMH for arXiv/PubMed, Project Gutenberg cache URLs, etc.)
• Per-item license filter (accepts ODC-BY, CC-BY-SA, CC-BY, CC0, MIT, Apache 2.0; rejects NC, ND, GPL/AGPL/LGPL per DESIGN.md §3)
• 768-character extracts with boundary truncation
• Manifest assembled with per-item sha256 cross-check
• Corpus hash and freeze date written to CORPUS_LOCK; once frozen, no in-place edits — corrections go to v2

Reproducibility: the build is fully scripted but uses external APIs that may be rate-limited or change over time. The frozen corpus.jsonl + manifest.yaml here is the authoritative artifact — re-running the build should produce the same content but may take days due to rate limits.

Considerations for use

In-scope:

• Generating activation sketches comparable with public v1 ckasketch sketches
• Cross-architecture model probing (the corpus is intentionally domain-mixed)
• Benchmarking activation-based RSA / CKA methods
• Per-item attribution lookup (use manifest.yaml)

Out of scope:

• Training data (this is intentionally a frozen, small, public probe set — not training material)
• Model fine-tuning (the per-item licenses don't all allow this; check NOTICES.md for any item you intend to redistribute)
• Re-extraction (corpus is intentionally frozen at extracted text; re-fetching from original sources may yield different content if the source has changed)

Mixing with vision/audio/multimodal: future ckasketch tracks (vision, audio, multimodal_vt) are documented in DESIGN.md §4 but not built yet. They'd ship in a separate v1/{vision,audio,multimodal_vt}/ subdirectory and would be independently corpus-hashed.

Citation

@misc{ckasketch-calibration-v1,
  author = {Jones, Marc},
  title = {ckasketch v1 calibration corpus (text)},
  year = {2026},
  publisher = {HuggingFace Hub},
  url = {https://huggingface.co/datasets/marcjon/ckasketch-calibration-v1},
  note = {Frozen 2026-05-17; corpus_hash sha256:cbd6a314d904842e1c5cda65eca146f8b7ddcd2027c6d0036f789a6d0a405c37; ODC-BY 1.0},
}

@software{ckasketch,
  author = {Jones, Marc},
  title = {ckasketch: CKA-based representational similarity sketches for ML models},
  url = {https://github.com/marctjones/ckasketch},
  year = {2026},
}

Cross-references

• ckasketch source repo: https://github.com/marctjones/ckasketch
• Sketches produced against this corpus: https://huggingface.co/datasets/marcjon/ckasketch-sketches
• DESIGN.md (full spec for tracks, licensing tiers, sketch format): https://github.com/marctjones/ckasketch/blob/main/ckasketch/calibration/DESIGN.md

License attribution

Compilation: ODC-BY 1.0. Each individual document retains its original source license — see NOTICES.md (line-by-line) and manifest.yaml (machine-readable). When redistributing or building derived works, attribute both the compilation (this dataset) and the underlying sources per their respective requirements.

Maintained by

@marcjon. Issues and corrections welcome at https://github.com/marctjones/ckasketch/issues. The corpus itself is frozen — any correction lands in a future v2 (with a new corpus_hash).