gspc-normalized / README.md
Nicholastempleman's picture
Add README explaining schema mapping
70f5c9c verified
|
Raw
History Blame Contribute Delete
2.19 kB
metadata
license: mit
tags:
  - csoai
  - gspc
  - governance
  - safety
  - evaluation
  - benchmark

csoai/gspc-normalized

Unified schema for the csoai gspc-* governance/safety evaluation corpus.

Schema

Every row has exactly these fields (one unified schema for all sources):

field type description
source string Original HF dataset id (csoai/gspc-*)
axis string Short axis id (gov, prv, asi, ...)
category string Original dataset's category/axis label (may be empty if source had none)
anchor string Original dataset's anchor/note/source/why field (may be empty)
prompt string The question/scenario text (the input to the model)
expected string The expected answer — string-scalar for classification, `
expected_is_list bool True if original gold was a list (multi-label source)
raw_keys string JSON dump of the original row's field names (debug aid)

Sources mapped

The underlying csoai/gspc-* datasets used six different field-name conventions for the prompt and gold fields. This dataset unifies them.

Source axis rows
csoai/gspc-agi agi 36
csoai/gspc-art5 art5 36
csoai/gspc-asi asi 33
csoai/gspc-det det 6
csoai/gspc-gov gov 237
csoai/gspc-mach mach 33
csoai/gspc-mcp mcp 35
csoai/gspc-oss oss 32
csoai/gspc-prv prv 32
csoai/gspc-swarm swarm 6
csoai/gspc-xr xr 32

Skipped

  • csoai/gspc-care — intentionally excluded. Its rows have only metadata (battery, timestamp, v2) with no prompt field. Including it would require a manual annotation pass to add prompts. The source dataset itself remains untouched on HF.

Provenance

Built 2026-08-08 by gspc_normalize.py after an external benchmark spray revealed the schema drift across the corpus. The unifying schema makes the gspc-* corpus usable by downstream benchmark runners, fine-tuners, and eval harnesses without per-dataset field-mapping code.

License

MIT. Original datasets were MIT per the csoai org convention.