Upload README.md with huggingface_hub

README.md

@@ -8,8 +8,10 @@ tags:
   - simulation
   - convergence-maps
   - des-y3
+  - lsst
+  - shape-noise
 size_categories:
-  - 10K<n<100K
+  - 100K<n<1M
 ---
 
 # Gower Street DES Y3 Lensing Tiles
@@ -18,33 +20,75 @@ Weak lensing convergence map tiles extracted from the [Gower Street](http://star
 
 ## Dataset Description
 
-Each sample contains a (4, H, W) convergence map tile covering ~3400 deg^2, corresponding to 4 DES Y3 MagLim tomographic bins. Tiles are extracted from equatorial HEALPix base faces after harmonic-space filtering and rotation for data augmentation.
+Each sample contains a (4, H, W) convergence map tile covering ~3400 deg², corresponding to 4 DES Y3 MagLim tomographic bins. Tiles are extracted from equatorial HEALPix base faces after harmonic-space filtering and rotation for data augmentation.
+
+The dataset includes **15 configurations** (5 angular scale cuts x 3 noise levels), with ~9400 tiles per configuration from 787 simulations (12 tiles per sim: 3 rotations x 4 equatorial tiles).
 
 ### Configurations
 
-The dataset is sharded by maximum multipole (`lmax`), which controls the angular scale content:
+Each configuration is identified by `lmax_{lmax}_{noise_level}` and stored in a separate data directory:
+
+| lmax | Tile size | Angular scales | nside | Noise levels |
+|------|-----------|---------------|-------|--------------|
+| 200  | 128x128   | > 0.9 deg     | 128   | noiseless, des_y3, lsst_y10 |
+| 400  | 256x256   | > 0.45 deg    | 256   | noiseless, des_y3, lsst_y10 |
+| 600  | 256x256   | > 0.3 deg     | 256   | noiseless, des_y3, lsst_y10 |
+| 800  | 512x512   | > 0.23 deg    | 512   | noiseless, des_y3, lsst_y10 |
+| 1000 | 512x512   | > 0.18 deg    | 512   | noiseless, des_y3, lsst_y10 |
+
+### Noise levels
+
+Shape noise arises from the intrinsic ellipticity dispersion of source galaxies. For a HEALPix pixel at resolution nside, the noise standard deviation per pixel per tomographic bin is:
+
+```
+sigma_pix = sigma_e / sqrt(2 * n_eff * A_pix)
+```
+
+where `sigma_e` is the per-component intrinsic ellipticity dispersion, `n_eff` is the effective galaxy number density (in sr⁻¹), and `A_pix = 4pi / N_pix` is the pixel solid angle. The factor of 2 accounts for two ellipticity components. Noise is Gaussian and independent per pixel.
+
+Shape noise is added to the full-sky nside=1024 convergence map **before** harmonic filtering, so the noise is band-limited consistently with the signal. For a given (sim_id, noise_level), the same noise realization is shared across all lmax cuts and orientations. RNG seed: `sim_id * 1000 + noise_level_index`.
 
-| Config | lmax | Tile size | Angular scales |
-|--------|------|-----------|---------------|
-| `lmax_200` | 200 | 128x128 | > 0.9 deg |
-| `lmax_400` | 400 | 256x256 | > 0.45 deg |
-| `lmax_600` | 600 | 256x256 | > 0.3 deg |
-| `lmax_800` | 800 | 512x512 | > 0.23 deg |
-| `lmax_1000` | 1000 | 512x512 | > 0.18 deg |
+#### `noiseless`
+
+No shape noise added. Pure signal from the Born-approximation raytracing.
+
+#### `des_y3` — DES Year 3 ([Amon et al. 2022](https://arxiv.org/abs/2105.13543), Table 1)
+
+Per-bin effective number density and intrinsic ellipticity dispersion from the DES Y3 MagLim sample:
+
+| Bin | n_eff (arcmin⁻²) | sigma_e |
+|-----|-------------------|---------|
+| 0   | 1.476             | 0.243   |
+| 1   | 1.479             | 0.262   |
+| 2   | 1.484             | 0.259   |
+| 3   | 1.461             | 0.301   |
+
+#### `lsst_y10` — LSST Year 10 ([DESC SRD](https://arxiv.org/abs/1809.01669))
+
+| Bin | n_eff (arcmin⁻²) | sigma_e |
+|-----|-------------------|---------|
+| 0-3 | 6.75              | 0.26    |
+
+Total n_eff = 27 arcmin⁻² split uniformly across 4 bins to match the DES tomographic structure.
 
 ### Loading
 
 ```python
 from datasets import load_dataset
 
-# Load a specific lmax configuration
-ds = load_dataset("EiffL/GowerStreetDESY3", data_dir="data/lmax_600")
+# Load a specific (lmax, noise_level) configuration
+ds = load_dataset("EiffL/GowerStreetDESY3", data_dir="data/lmax_600_des_y3")
 
-# Access a sample
 sample = ds["train"][0]
-kappa = sample["kappa"]       # (4, 256, 256) convergence map
-omega_m = sample["Omega_m"]   # Matter density parameter
-s8 = sample["S8"]             # S8 = sigma_8 * sqrt(Omega_m / 0.3)
+kappa = sample["kappa"]           # (4, 256, 256) convergence map
+omega_m = sample["Omega_m"]       # Matter density parameter
+noise = sample["noise_level"]     # "des_y3"
+
+# Load noiseless version at same angular scale
+ds_clean = load_dataset("EiffL/GowerStreetDESY3", data_dir="data/lmax_600_noiseless")
+
+# Load LSST-depth version
+ds_lsst = load_dataset("EiffL/GowerStreetDESY3", data_dir="data/lmax_600_lsst_y10")
 ```
 
 ### Fields
@@ -55,6 +99,7 @@ s8 = sample["S8"]             # S8 = sigma_8 * sqrt(Omega_m / 0.3)
 | `sim_id` | int | Gower Street simulation ID (1-791) |
 | `orientation_id` | int | Rotation orientation (0-2) |
 | `tile_id` | int | Equatorial tile index (0-3) |
+| `noise_level` | string | Noise level: "noiseless", "des_y3", or "lsst_y10" |
 | `Omega_m` | float | Matter density parameter |
 | `sigma_8` | float | RMS density fluctuation amplitude |
 | `S8` | float | S8 = sigma_8 * sqrt(Omega_m / 0.3) |
@@ -67,10 +112,10 @@ s8 = sample["S8"]             # S8 = sigma_8 * sqrt(Omega_m / 0.3)
 ## Pipeline
 
 1. **N-body simulations**: Gower Street suite (791 simulations with varying cosmological parameters)
-2. **Raytracing**: Born-approximation lensing through shell outputs, weighted by DES Y3 MagLim n(z) distributions (4 tomographic bins)
-3. **Harmonic filtering**: `map2alm(lmax)` -> `rotate_alm(euler)` -> `alm2map(nside_down)` ensures all tiles see identical harmonic-space processing
-4. **Tile extraction**: 3 fixed rotations x 4 equatorial HEALPix base tiles = 12 tiles per simulation
-5. **Downsampling**: nside matched to lmax (Nyquist criterion)
+2. **Raytracing**: Born-approximation lensing through particle lightcone shells (nside=2048 input, nside=1024 output), weighted by DES Y3 MagLim n(z) distributions (4 tomographic bins)
+3. **Shape noise injection**: Gaussian noise added per pixel at nside=1024, calibrated to DES Y3 or LSST Y10 survey depth
+4. **Harmonic filtering**: `map2alm(lmax)` -> `rotate_alm(euler)` -> `alm2map(nside_down)` ensures all tiles see identical harmonic-space processing
+5. **Tile extraction**: 3 fixed rotations x 4 equatorial HEALPix base tiles = 12 tiles per simulation per configuration
 
 ### Rotations