artifacts

Paper artifacts

The scored results behind the paper's figures. These are the merged CSVs the plotting scripts read; nothing here needs a GPU, a model server or a precompute store to redraw.

What is here

One directory per experiment, in the layout the plotting scripts expect (<task>/merged/<benchmark>/<split>/<producer>.csv), so this folder can be passed straight to --results-root.

task	producer	CSV	compares
base01	baselines	baselines.csv	Stretto against Lotus and Abacus
samp01	sample_size	sample_size.csv	cost and accuracy against profiling budget
ops01	operator_count	operator_count.csv	cost and accuracy against operator set and storage
abl01	ablation	ablation.csv	Stretto, minus compression, minus the optimizer
mode01	baselines	baselines.csv	global against per-step optimization, over the full operator set at n=150
ref01	label_reference	label_reference_{gold,silver}_metrics.csv	model-derived against human labels

The first five cover five random benchmarks each — artwork_random_medium, rotowire_random, movie_random_huge, email_random, ecommerce_random_large. ref01 covers the five curated ones, which are the benchmarks carrying per-tuple ground truth.

single_operator/ holds the single-operator KV cache results behind Figure 6 (see its README).

kvops/kvop01/ holds the KV-operator sweep behind the footprint-vs-speedup figure (kvop01/merged/<benchmark>/dev/kv_operator.csv): the uncompressed default suite plus one KV-compressed operator per state, against the suite without it.

Redrawing the figures

# One task. Every figure family, both pooling rules, all dataset panels.
python scripts/plot_sweep.py --experiment base01 --results-root artifacts

# ref01 has its own script
python scripts/plot_label_reference.py --output-dirs artifacts/ref01/merged --figures strip

# The KV-operator sweep (reads artifacts/kvops by default)
python scripts/plot_kvop.py kvop01 --target-ratio 0.7 0.9 --compact --metric total --aggregate geomean

Figures land in the task's own directory unless --figure-dir (or --out-dir, for plot_label_reference.py) says otherwise. See "The paper's figures" in the README at the repository root for the exact command behind each figure.