README.md

---
license: mit
task_categories:
- text-classification
language:
- en
tags:
- system-one
- typed-decisions
- synthetic
size_categories:
- 10K<n<100K
pretty_name: systemone-lite-general
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: test
    path: data/test-*
  - split: test_hard
    path: data/test_hard-*
  - split: full
    path: data/full-*
---

# systemone-lite-general

Synthetic typed-decision rows for
[`systemone-lite`](https://github.com/fritzprix/systemone-lite)
(letter-alias `choice` labels for causal LM SFT).

Not affiliated with TypeSafe AI / Jev. Labels are rule-based, not human prefs.

## Splits

| Split | Rows | Notes |
|------|------|------|
| `train` | 32 400 | Stratified mix of 3 gyms |
| `test` | 3 600 | iid held-out by task |
| `test_hard` | 5 400 | layout / paraphrase / option-subset shift |
| `full` | 36 000 | train + iid test |

## Gyms

- **TicketDungeon**: `ticket.route`, `ticket.needs_human`, `ticket.urgency`
- **ResourceAllocator**: `alloc.fund_next`, `alloc.can_fund_all`, `alloc.pressure`
- **DebateJudge**: `debate.winner`, `debate.enough_evidence`, `debate.confidence`

## Downstream numbers (same repo)

Fine-tune → [`dwidlee/systemone-lite-0.5b`](https://huggingface.co/dwidlee/systemone-lite-0.5b):

| Split | Base 0.5B | SFT |
|---|---:|---:|
| `test` (n=3600) | 0.439 | 0.679 |
| `test_hard` (n=5400) | 0.427 | 0.652 |

## Load

```python
from datasets import load_dataset
ds = load_dataset("dwidlee/systemone-lite-general")
print(ds)
```