Datasets:
File size: 3,096 Bytes
232313d 3328ee4 d25cd8a 232313d 35af102 3328ee4 35af102 3328ee4 35af102 3328ee4 35af102 3328ee4 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 | # Jev Skills code
Standalone Python (3.10+). Reading parquet needs `pandas` and `pyarrow`; rebuilding `rules7.py` also needs `huggingface_hub`.
No `jevswarm` imports.
| file | what |
|---|---|
| `score.py` | the scorer (`load_items`, `score`, `evaluate`, oracle / uniform baselines) |
| `spatial_bench.py` | builds JevSpatial v1 (`games_spatial/*` 1-5), a few seconds, no network |
| `play_bench.py` | closed-loop game simulations (Snake, Dino, Breakout, Whac-A-Mole) driven by any decider: `--decider python --decider-target module:factory`, `--decider http --url ...` (any server speaking the System One wire format, `POST /v1/systemone` with `{"model", "state", "questions"}`), `oracle`, `random` |
| `rules7.py` | builds Jev Rules-7 (downloads the pinned ShARC dev and patent test files), uses `drop_keys.json` (rule texts excluded from the ShARC selection) |
| `browser_bench.py` | builds `agents/browser` |
| `bench_common.py`, `rows.py` | helpers; `rows.py` maps generator skills to `category/skill` and licence ids |
| `docs/` | notes on JevSpatial and Rules-7 |
## Rebuild the items
```bash
python spatial_bench.py --out out/ # jevspatial_v1_items.json.gz (2,040 items)
python rules7.py --out out/ # rules7_v1_items.json.gz (3,440 items; needs network)
python -c "import browser_bench as b; items, info = b.build(0); print(len(items))" # 600 items
```
The output is deterministic and byte-identical to the published rows for these three parts (checked when the dataset was
built). The other skills (tools, agent_action, chess, small_games, safety, judging, grounding) are shipped as data only;
their sources and pinned revisions are in the dataset card.
## Score a model
```python
import score
def decide(state, questions):
# call your model; return one answer per question key
return {k: ({"noul": 0.5} if q["type"] == "noul" else
{"probabilities": {o: 1 / len(q["criteria"]) for o in q["criteria"]}})
for k, q in questions.items()}
items = score.load_items("../data/rules.parquet") # or parquet / jsonl / json.gz
print(score.evaluate(decide, items)["pooled_by_category"])
```
For `score` questions return probabilities keyed by the level strings in `criteria` (or their indices).
`python score.py ../data/rules.parquet --baseline oracle` prints 1.0 for every skill.
## Closed-loop games with your own decider
`play_bench.py` plays Snake, Dino, Breakout and Whac-A-Mole with any decider. A Python hook, `my_decider.py`:
```python
def make_decider():
def decide(state, questions): # uniform guesser
return {k: {"noul": 0.5} for k in questions}
return decide
```
`python play_bench.py --decider python --decider-target my_decider:make_decider --decisions 100`. The factory may also return an
object with a `.decide(state, questions)` method. Any server speaking the System One wire format works with
`--decider http --url http://HOST:PORT` (it posts `{"model", "state", "questions"}` to `/v1/systemone`); `--decider oracle`
(always the best move) and `--decider random` are references.
|