---
license: apache-2.0
language:
- en
pretty_name: skycmd secrets jev
task_categories:
- text-classification
tags:
- skycmd
- typed-decisions
- system-one
configs:
- config_name: secret_bench
  data_files: data/secret_bench.parquet
- config_name: secret_synth
  data_files: data/secret_synth.parquet
---

# skycmd secrets jev

Short texts (.env, code, configs, curl, logs, diffs) with four questions each: is a usable secret exposed, which service issued it, what kind of value it is, on which line. Every key is generated: it follows its provider's public format with a random body, so none is a real key. `secret_bench` uses four file types the training set never shows.

Part of [skycmd](https://github.com/maxgfr/skycmd), tiny decision models trained from scratch in Rust that run in the
browser. Written by `sky-data secret-jev (seed 1234)` at commit [`6cbcce8c`](https://github.com/maxgfr/skycmd/commit/6cbcce8cbbbf27e8e952e61b6cb1ed03ee1fca58).
Licences of every set: [DATA_LICENSES.md](https://github.com/maxgfr/skycmd/blob/main/DATA_LICENSES.md).

Each row is one typed decision question (docs/SPEC.md):

| field | type | meaning |
|---|---|---|
| `src` | string | source |
| `t` | string | `choice`, `score` or `noul` |
| `ins` | string | the question |
| `opts` | list of strings | the options, rendered as text |
| `state` | string | the state the question is about |
| `target` | list of floats | a probability per option (soft or one-hot) |
| `split` | string | `train`, `val` or `test` |

## Licence

Apache-2.0: generated by skycmd's own code, no third-party data.

## Files

| file | rows |
|---|---:|
| `data/secret_bench.parquet` | 7,984 |
| `data/secret_synth.parquet` | 193,258 |

```python
from datasets import load_dataset
ds = load_dataset("maxgfr/skycmd-secrets", "secret_bench")
```

skycmd wrote no row with a language model: the targets come from the upstream labels (some upstream
sets carry their own teacher soft labels, see their cards), the games' oracles or our generators.