Datasets:

fiktor
/

FIM-Estimation

@@ -15,6 +15,11 @@ configs:
     path: isnnn_400/seed_01/d_train.parquet
   - split: test
     path: isnnn_400/seed_01/d_test.parquet
 - config_name: isnnn_400.seed01.gt_fim
   data_files:
   - split: test
@@ -111,15 +116,19 @@ configs:
     path: isnnn_400/seed_10/gt_fim.parquet
 ---
-# FIM-Estimation
 In a FIM-Estimation task, the input consists of the following components:
 1. A dataset of the form $\mathcal{D}_ {train} = \{(\mathbf{\lambda}_ i, x_ i)\}_ {i=1}^{\left|\mathcal{D}\right|}$, where $\vlambda_ i$ is a point in the parameter space and $x_ i$ is a sample.
 2. The structure of the samples $x_ i$. For example, if $x_ i\in\{0,1\}^n$ is a bitstring representing a measurement of a quantum system on a lattice with $n$ sites, then the said structure is the lattice and the correspondence between the bitstring bits and the lattice sites.
 3. Optionally, additional information about the statistical manifold. This could involve, e.g., the symmetries of the system used to generate the statistical manifold.
-The task is to estimate Fisher Information Metric as a matrix-valued function of parameter $\mathbf{\lambda}$.
 We (TODO:plan to) present datasets corresponding to 6 statistical manifolds: `XXZ300_Z`, `FIL24`, etc.
 The data is stored as follows:
 `data/xxz300_z` contains the data for `XXZ300_Z` datasets. Within that directory, there are subdirectories of the form `seed_??` each containing a dataset describing the same statistical manifold, but generated using a different seed (e.g. `seed_05`). We have `seed_05/d_train.parquet`, which is the dataset as described in #1 above. `seed_05/d_test.parquet` is a hold-out dataset which should not be used in estimation of $\mathbf{\lambda}$ (not even for hyperparameter tuning).

     path: isnnn_400/seed_01/d_train.parquet
   - split: test
     path: isnnn_400/seed_01/d_test.parquet
+  description: >-
+    Samples for IsNNN400 statistical manifold generated with seed=1.
+    `sample` column is formatted as bytearray of length 50 bytes (400 bits)
+    and represents a spin configuration of 20x20 lattice
+    with bit 0 corresponding to Z=+1 and bit 1 corresponding to Z=-1.
 - config_name: isnnn_400.seed01.gt_fim
   data_files:
   - split: test
     path: isnnn_400/seed_10/gt_fim.parquet
 ---
+# Dataset Card for FIM-Estimation
+## Dataset Summary
 In a FIM-Estimation task, the input consists of the following components:
 1. A dataset of the form $\mathcal{D}_ {train} = \{(\mathbf{\lambda}_ i, x_ i)\}_ {i=1}^{\left|\mathcal{D}\right|}$, where $\vlambda_ i$ is a point in the parameter space and $x_ i$ is a sample.
 2. The structure of the samples $x_ i$. For example, if $x_ i\in\{0,1\}^n$ is a bitstring representing a measurement of a quantum system on a lattice with $n$ sites, then the said structure is the lattice and the correspondence between the bitstring bits and the lattice sites.
 3. Optionally, additional information about the statistical manifold. This could involve, e.g., the symmetries of the system used to generate the statistical manifold.
+The task is to estimate Fisher Information Metric (FIM) as a matrix-valued function of parameter $\mathbf{\lambda}$.
+In this dataset we present multiple such tasks (parameterised by task name and random seed). E.g. in order to solve the IsNNN400 task with seed=4 one would download config `isnnn_400.seed04`, split `train`, estimate the FIM, and compare the results with `isnnn_400.seed04.gt_fim`.
+## Additional information
 We (TODO:plan to) present datasets corresponding to 6 statistical manifolds: `XXZ300_Z`, `FIL24`, etc.
 The data is stored as follows:
 `data/xxz300_z` contains the data for `XXZ300_Z` datasets. Within that directory, there are subdirectories of the form `seed_??` each containing a dataset describing the same statistical manifold, but generated using a different seed (e.g. `seed_05`). We have `seed_05/d_train.parquet`, which is the dataset as described in #1 above. `seed_05/d_test.parquet` is a hold-out dataset which should not be used in estimation of $\mathbf{\lambda}$ (not even for hyperparameter tuning).