Download arrowspace-tests/controlled/CALIBRATION.md from mec-is/arrowspace-datasets-tests: direct link, hf CLI and curl.
- Browser
- Download file 4.68 kB
-
https://huggingface.co/datasets/mec-is/arrowspace-datasets-tests/resolve/main/arrowspace-tests/controlled/CALIBRATION.md
- Command line
-
hf download hf://datasets/mec-is/arrowspace-datasets-tests/arrowspace-tests/controlled/CALIBRATION.md
-
curl -L -o CALIBRATION.md https://huggingface.co/datasets/mec-is/arrowspace-datasets-tests/resolve/main/arrowspace-tests/controlled/CALIBRATION.md
CALIBRATION — synthetic dataset for the tests/test_0_*.py scripts
Empirical findings from calibrating a shared controlled dataset against
arrowspace==0.27.2 (pyarrowspace build, CPython 3.13, macOS arm64).
Final dataset: eigenmaps_controlled.parquet — 1000 items x 128 dims,
3 orthonormal clusters of increasing sparsity, with a cluster label column.
Regenerate with uv run python tests/data/make_datasets.py (seed 11).
1. Why the old tests failed
| Test | Failure | Root cause |
|---|---|---|
test_0_0, test_0_5 |
assert(hits[0][0] == 4) at tau<=0.9 |
5×24 dataset of near-identical rows → lambda spread is numerical noise, so the spectral term never reorders results. Expectations encoded 0.26-era lambdas; file last touched at v0.26.0. |
test_0_9 |
assert(len(hits) == 3) |
Contradicted build params topk: 5; 0.27.x correctly returns 5 hits. |
All three failed identically on 0.27.0 and 0.27.2 — pre-existing staleness,
not a regression from the version bump. The CI suite (pytest-tests/) passes
66/66 on 0.27.2.
2. 0.27.x scoring semantics (measured)
search(q, gl, tau)score =tau·cos(q, item) + (1−tau)·spec(item)(blend model confirmed exactly: predicted vs actual orders matched on every probed config).specis notλidirectly: the query is placed into the graph and gets its own lambdaλq(prepare_query_item);specmeasures spectral closeness betweenλqand the item'sλi. The shape is nonlinear — do not model it, measure it:search(q, gl, 0.0, k=n)(tau=0 ⇒ score IS spec).q = items[j] * scaleis cosine-invariant (cos(q, items[j]) = 1.0for anyscale > 0), butλqmoves withscale— the only query-side tuning knob.q = items[j]exactly (scale 1.0) raisesValueError: degenerate lambda (raw=0.000000). Always scale the query (e.g. 1.05–1.2).aspace.lambdas()are min-max normalized: exactly one item shows0.0(the floor). Harmless — the degenerate gate only fires for a query's own lambda. Quality gate:count(λ < 1e-12) <= 1for these datasets.eps=0.05(the old tests' value) leaves high-dim unit-norm data near-disconnected. These datasets useeps=0.5, sigma=0.5.- Dense noise σ drives λ per item, roughly quadratically; sparse 2-dim spikes give far less λ per L² (L²=0.078 → λ=0.032 vs dense σ=0.02 curve).
- Lifting an item past a cos-1.0 self-anchor at tau=0.9 needs
0.1·Δspec > 0.9·(1−cos)— a 9× barrier; naive twin-item designs sit right on it (see sweep log). Cluster-level invariants are far more robust.
3. Final dataset design (1000 x 128, seed 11)
3 clusters with orthonormal centroids (QR of random Gaussian), unit-norm rows:
| cluster | items | noise σ | mean λ (build_and_store) |
|---|---|---|---|
| 0 (dense) | 400 | 0.02 | 0.0451 |
| 1 (medium) | 350 | 0.05 | 0.1109 |
| 2 (sparse) | 250 | 0.12 | 0.1372 |
Graph params for all variants: eps=0.5, k=12, p=2.0, sigma=0.5;
topk=3 (tests 0_0, 0_5) / topk=5 (test 0_9).
Query: items[0] * 1.2 (dense-cluster member, scaled — never exact).
Verified invariants (all three builder variants)
build_and_storeandbuild_fullproduce identical results.- Sampled builder must use
with_sampling("simple", 1.0)for a small dataset (0.6 changes the graph; 1.0 keeps all items). - Cluster containment: every hit at every
tau ∈ {1.0, 0.9, 0.6, 0.55}is from the query's own cluster. - Self item ranks first at every tau.
- Mean λ increases with cluster sparsity:
λ̄₀ < λ̄₁ < λ̄₂(the spectral statistic tracks structural density). - Exactly one min-max zero λ.
- Bonus: within the dense cluster, tau ≤ 0.6 reorders the runner-ups (spectral blending active) — observed, not asserted (brittle).
4. Sweep log (calibration journey, kept for reference)
| Stage | Dataset | Outcome |
|---|---|---|
| probe (d=24) | 5 handcrafted rows | spec ≈ λ-closeness to query's λ discovered; exact-self query is degenerate |
| v2–v7 (d=60) | 5 twin/filler rows | 9× cosine-economics barrier; min-max λ floor; fillers hijack at scale ≥ 1.6; linear spec model insufficient at large offsets |
| measurement | d=60 twins | blend model tau·cos + (1−tau)·spec CONFIRMED via tau=0 spec extraction; winner found for [2,1,4]→[1,2,4] |
| final | 1000×128 clusters | cluster-level invariants: robust on every builder variant; adopted |
5. File inventory
eigenmaps_controlled.parquet— the shared dataset (featuresf0..f127+cluster).make_datasets.py— deterministic generator.CALIBRATION.md— this document.