README.md

---
configs:
- config_name: numeric_copy_increment
  data_files:
  - split: validation
    path: numeric_copy_increment/validation-*.parquet
- config_name: numeric_compare_sort
  data_files:
  - split: validation
    path: numeric_compare_sort/validation-*.parquet
- config_name: numeric_range_checksum_base
  data_files:
  - split: validation
    path: numeric_range_checksum_base/validation-*.parquet
- config_name: tabular_tsv_csv
  data_files:
  - split: validation
    path: tabular_tsv_csv/validation-*.parquet
- config_name: network_ip_port_rows
  data_files:
  - split: validation
    path: network_ip_port_rows/validation-*.parquet
- config_name: svg_path_numeric_blobs
  data_files:
  - split: validation
    path: svg_path_numeric_blobs/validation-*.parquet
- config_name: json_numeric_arrays
  data_files:
  - split: validation
    path: json_numeric_arrays/validation-*.parquet
- config_name: mmcif_coordinate_tables
  data_files:
  - split: validation
    path: mmcif_coordinate_tables/validation-*.parquet
- config_name: format_preserving_transforms
  data_files:
  - split: validation
    path: format_preserving_transforms/validation-*.parquet
- config_name: chunk_boundary_stress
  data_files:
  - split: validation
    path: chunk_boundary_stress/validation-*.parquet
---

# Synthetic numeric/structured-format PPL stress set

This dataset contains deterministic supervised PPL examples for probing numeric tokenization, local
format continuation, and structured numeric records in model perplexity-gap dashboards.

Each config has a `validation` split with `input` and `target` columns. The `input` field includes
three demonstrations and ends with `Assistant:\n`; the `target` field is the exact continuation to
score with target-only loss.

Repo: `marin-community/synth-numeric-format-ppl`
Seed: `50935010`
Rows per config: `1000`

## Configs

- `numeric_copy_increment`: 1000 validation examples
- `numeric_compare_sort`: 1000 validation examples
- `numeric_range_checksum_base`: 1000 validation examples
- `tabular_tsv_csv`: 1000 validation examples
- `network_ip_port_rows`: 1000 validation examples
- `svg_path_numeric_blobs`: 1000 validation examples
- `json_numeric_arrays`: 1000 validation examples
- `mmcif_coordinate_tables`: 1000 validation examples
- `format_preserving_transforms`: 1000 validation examples
- `chunk_boundary_stress`: 1000 validation examples

## Columns

- `id`, `subset`, `task`, `seed`
- `format`, `delimiter`, `difficulty`, `num_values`
- `input`, `target`
- `metadata_json`

The examples are generated, synthetic, and intended for evaluation/diagnostics rather than model
training claims.