tensordex/tensordex-ae-cache

TensorDex — Artifact Evaluation Cache

Data companion for the TensorDex SOSP artifact (code repo → ae/). It holds the pre-computed results and the raw tensors needed to reproduce the paper's figures and to verify them from scratch, so reviewers don't re-run the full 40 TB pipeline.

The code repo's ae/download_cache.py pulls this into ae/cache/.

Contents

Path	Size	Role
results.db	5.2 GB	the 11.4 M-pair compression cache (every figure's numbers)
sample_blobs/<xx>/<yy>/<id>.safetensors	~2 GB	raw tensor blobs for the Tier-1 sample (content-addressed)
sample_pairs.tsv	—	the (target, base) pairs the sample covers
data/tensordb_s3/metadata.db	567 MB	slim: tensor sizes + model→tensor map (for reduction charts)
model_hub_crawl/	~127 MB	Hugging Face crawl stats (Fig 2, Fig 4)
tests/output/, compression_data/, model_level_reduction/	~500 MB	baseline plans & traces the chart modules read

results.db is a self-contained SQLite file (WAL folded in), slimmed to exactly the columns the AE charts and verification scripts read (original row ids are preserved, so subsampled charts reproduce the paper's figures byte-for-byte). Blobs are keyed by XXH3-128 of the raw tensor bytes — id == hash(bytes); that is what Tier-1 verify_sample.py re-checks.

results.db column guide

One row per compressed (target, base) tensor pair of the trace:

Column(s)	Meaning
target_id, base_id	XXH3-128 content hashes of the two tensors (= blob ids)
param_name, shape, bytes_in	tensor identity and raw size
bcs_dist	TensorSketch (BCS) distance between the pair — TensorPred's input
tratio / tbytes_out	TensorX delta codec (TensorDex-TX, 65.1 %) — re-derived bit-exact by make verify
fratio / fbytes_out	FM++ delta codec (the 70.5 % headline) — re-derived bit-exact after make ae-fmpp
ratio / bytes_out	BitX (ZipLLM's delta codec) baseline
aratio / abytes_out, zratio	dense per-pair delta ratios computed during development — the fitting cache TensorPred is trained and evaluated on (Fig 13; 5.77 M pairs vs ~1 M for FM++); zratio is its zstd fallback for tiny tensors, also used by the entropy/CDF analyses
pred_ratio	TensorPred's stored prediction of aratio — make verify-predict re-fits the model from scratch and recovers this column to ~1e-4
timestamp, ttimestamp	row provenance (trace ordering; legacy-row reporting in verify)

Use

# in the code repo
python ae/download_cache.py --repo <this-dataset-id>
make figures      # Tier 0 — re-plot every figure
make verify       # Tier 1 — re-derive a random sample from these blobs

Provenance

Generated by the code repo's authoring tools: ae/stage_data.py (chart inputs

• slim metadata.db), ae/build_sample_bundle.py (the blob sample). Regenerate a larger sample with --budget-gb N.