borisbolliet/cmbemu-competition-v1

---
license: mit
task_categories:
- other
tags:
- cosmology
- CMB
- emulator
- benchmark
- power-spectrum
size_categories:
- 10K<n<100K
---

# cmbemu competition v1 — CMB power-spectrum emulator benchmark

Pre-generated CMB angular power spectra for a Latin-hypercube sweep over a 6-parameter cosmology box, produced by [`classy_szfast`](https://github.com/CLASS-SZ/classy_szfast) with `cosmo_model=6` (ede-v2 emulator stack, near-LCDM in this configuration). Intended as the official data distribution for the `cmbemu` emulator-competition package.

## Contents

| File | N | Size (float32) | Purpose |
|-------------------|:------:|:--------------:|---------|
| `train.npz` | 50 000 | ~3.7 GB | Training set for emulator authors |
| `test.npz` | 5 000 | ~0.4 GB | Scoring test set |
| `train_small.npz` | 5 000 | ~0.4 GB | Dev-loop training set |
| `test_small.npz` | 500 | ~40 MB | Dev-loop test set |

All four files share the same schema.

## Schema

Each `.npz` contains:

| Key | Shape | dtype | Description |
|---------------------|------------------------|:--------:|-------------|
| `params` | (N, 6) | float32 | LHC points in physical units |
| `param_names` | (6,) | str | Order of the 6 cosmology parameters |
| `tt` | (N, 6001) | float32 | C_ℓ^TT for ℓ ∈ [0, 6000] |
| `te` | (N, 6001) | float32 | C_ℓ^TE |
| `ee` | (N, 6001) | float32 | C_ℓ^EE |
| `pp` | (N, 3001) | float32 | C_ℓ^{φφ} for ℓ ∈ [0, 3000] |
| `box_lo`, `box_hi` | (6,) | float32 | Competition box bounds |
| `lmax_cmb`, `lmax_pp` | () | int32 | 6000, 3000 |
| `cosmo_model` | () | int32 | 6 (= `ede-v2` in classy_szfast) |
| `classy_sz_version` | () | str | Exact generator version |
| `seed` | () | int64 | RNG seed used for the LHC |

All ℓ-indexed arrays start at ℓ=0. The monopole/dipole (ℓ=0, 1) are ignored by the likelihood — slice from ℓ=2 when scoring.

## Parameter box

The 6D competition prior covers roughly ±10σ–30σ around Planck 2018, strictly inside the `classy_szfast` ede-v2 emulator validity region:

| Parameter | Min | Max |
|-----------------|:------:|:------:|
| ω_b | 0.020 | 0.025 |
| ω_cdm | 0.09 | 0.15 |
| H₀ | 55 | 85 |
| τ | 0.03 | 0.10 |
| ln 10¹⁰ A_s | 2.7 | 3.3 |
| n_s | 0.92 | 1.02 |

## Quickstart

```python
import cmbemu as cec

# Downloads the .npz files on first use, cached under ~/.cache/cmbemu
train = cec.load_train()
test = cec.load_test()

# Score your emulator
class MyEmulator:
def predict(self, params: dict) -> dict:
...

result = cec.get_score(MyEmulator())
print(result["combined_S"], result["mae_total"]["mae"])
```

See the [cmbemu repository](https://github.com/borisbolliet/cmbemu) for the full scoring definition and competition rules.

## Generation reproducibility

Regenerate any subset locally:

```python
import cmbemu as cec
data = cec.generate_data(n=5000, seed=202605, save_to="test_reproduced.npz")
```

Seeds used here: `202604` for train, `202605` for test.

## License

MIT.