---
pretty_name: Samsoup GraphArch
configs:
- config_name: full
  data_files:
  - split: train
    path: full/train.jsonl
  - split: validation
    path: full/validation.jsonl
  - split: test
    path: full/test.jsonl
- config_name: subset_10k
  data_files:
  - split: train
    path: subset_10k/train.jsonl
  - split: validation
    path: subset_10k/validation.jsonl
  - split: test
    path: subset_10k/test.jsonl
---

# GraphArch

This dataset is the MO-RELISH graph architecture regression collection. Each
row contains serialized neural-network graph text and execution/benchmark
measurements from neural architecture search spaces.

## Configs And Splits

| Config | Train | Validation | Test |
| --- | ---: | ---: | ---: |
| `full` | 475,257 | 10,000 | 10,000 |
| `subset_10k` | 10,000 | 1,000 | 1,000 |

The `full` config keeps all strict-clean rows, with 10,000 validation rows and
10,000 test rows sampled deterministically and stratified by search space. The
`subset_10k` config is a deterministic 10,000/1,000/1,000 sample from the full
splits, also stratified by search space.

## Columns

Input columns:

- `source_text`: task-level context.
- `input_text`: serialized neural-network graph text.
- `reference_outputs`: empty list; this is a regression benchmark.
- `prompt_components.problem_context`: same task context as `source_text`.
- `prompt_components.input_to_evaluate_column`: points to `input_text`.
- `prompt_components.input_format`: serialized neural-network graph text.
- `prompt_components.search_space`: NAS search-space name.

Prediction targets:

- `targets.val_accuracy`: Validation accuracy reported for the neural architecture.
- `targets.flops`: Floating-point operation count for the architecture.

Retained measurements:

- `measurements.val_accuracy`: Validation accuracy reported for the neural architecture.
- `measurements.flops`: Floating-point operation count for the architecture.
- `measurements.params`: Number of trainable parameters in the architecture.

`params` is retained for analysis but is not a default prediction target because
it is usually deterministic from the architecture.

The original zero-cost-proxy metadata is not included in rows because it may
leak target values such as FLOPs, parameter count, or validation accuracy.

## Loading

```python
from datasets import load_dataset

full_ds = load_dataset("Samsoup/GraphArch", "full")
small_ds = load_dataset("Samsoup/GraphArch", "subset_10k")
```