Datasets:
metadata
license: other
license_name: per-source
license_link: https://github.com/maxgfr/skycmd/blob/main/DATA_LICENSES.md
language:
- en
- fr
pretty_name: skycmd typed decisions
task_categories:
- text-classification
tags:
- skycmd
- typed-decisions
- system-one
configs:
- config_name: allocine
data_files: data/allocine.parquet
- config_name: arc
data_files: data/arc.parquet
- config_name: assistant_decisions
data_files: data/assistant_decisions.parquet
- config_name: banking77
data_files: data/banking77.parquet
- config_name: boolq
data_files: data/boolq.parquet
- config_name: calibrated_typed
data_files: data/calibrated_typed.parquet
- config_name: commonsense_qa
data_files: data/commonsense_qa.parquet
- config_name: hellaswag
data_files: data/hellaswag.parquet
- config_name: helmo_synth
data_files: data/helmo_synth.parquet
- config_name: helpsteer2
data_files: data/helpsteer2.parquet
- config_name: mnli
data_files: data/mnli.parquet
- config_name: neuro_laya
data_files: data/neuro_laya.parquet
- config_name: notam
data_files: data/notam.parquet
- config_name: open_jev
data_files: data/open_jev.parquet
- config_name: pawsx_fr
data_files: data/pawsx_fr.parquet
- config_name: pngwn_typed
data_files: data/pngwn_typed.parquet
- config_name: procedural_typed
data_files: data/procedural_typed.parquet
- config_name: qasc
data_files: data/qasc.parquet
- config_name: refusal
data_files: data/refusal.parquet
- config_name: snli
data_files: data/snli.parquet
- config_name: swarm_laya
data_files: data/swarm_laya.parquet
- config_name: tasksource_jev
data_files: data/tasksource_jev.parquet
- config_name: tasksource_synth
data_files: data/tasksource_synth.parquet
- config_name: tetris_laya
data_files: data/tetris_laya.parquet
- config_name: trading_laya
data_files: data/trading_laya.parquet
- config_name: typed_synth
data_files: data/typed_synth.parquet
- config_name: when2call
data_files: data/when2call.parquet
skycmd typed decisions
Typed decision questions (choice / score / noul with a probability per option) converted from public datasets with a permissive or share-alike licence, the decision data of the skycmd s3 to s6 models.
Part of skycmd, tiny decision models trained from scratch in Rust that run in the
browser. Written by sky-data fetch-decisions data/decisions (seed 1234, default caps) at commit 6cbcce8c.
Licences of every set: 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 |
Sources and licences
Each config is one source and keeps its source's licence (share-alike sources stay share-alike). Attribution goes to the upstream authors; the conversion is ours.
| config | upstream | licence | how it is used | rows |
|---|---|---|---|---|
allocine |
tblard/allocine | mit | positive review noul in French | 30,000 |
arc |
allenai/ai2_arc | cc-by-sa-4.0 | choice A..E | 7,784 |
assistant_decisions |
kgrozdanovski/assistant-decisions | apache-2.0 | gold probability distributions | 15,455 |
banking77 |
PolyAI/banking77 | cc-by-4.0 | gold intent + 7 distractors | 13,083 |
boolq |
google/boolq | cc-by-sa-3.0 | noul over the passage | 12,697 |
calibrated_typed |
m-lagnajit/calibrated-typed-decisions | cc-by-4.0 | permissive rows, calibrated teacher_probs | 18,794 |
commonsense_qa |
tau/commonsense_qa | mit | choice A..E | 10,962 |
hellaswag |
Rowan/hellaswag | mit | choice A..D over the endings, 30k | 30,000 |
helmo_synth |
helmo/synthetic-typed-decisions | mit | gold labels | 6,574 |
helpsteer2 |
nvidia/HelpSteer2 | cc-by-4.0 | score questions (0-4) per attribute | 60,000 |
mnli |
nyu-mll/glue (mnli) | other (MultiNLI: OANC terms, some fiction CC-BY-SA-3.0 / CC-BY-3.0) | entailment / neutral / contradiction, balanced 60k | 60,000 |
neuro_laya |
Aditharavind/laya-neuro-decisions (from STEW, Lim et al. 2018) | cc-by-4.0 | EEG workload noul | 16,456 |
notam |
AirsideLabs/notam-typed-decisions | cc-by-4.0 | evaluation only (split test) | 9,051 |
open_jev |
ZefanCai/Open-Jev (browser-drone-expansion-v1-redistributable) | cc0-1.0 | drone-control kept whole, 30k others | 47,800 |
pawsx_fr |
google-research-datasets/paws-x (fr) | other (PAWS-X: free for any purpose, acknowledgement of Google appreciated) | paraphrase noul in French | 48,864 |
pngwn_typed |
pngwn/typed-decisions | cc-by-sa-4.0 | GPT-2 prompts decoded to text, gold_probs as soft targets | 51,019 |
procedural_typed |
tasksource/procedural-typed-decisions | apache-2.0 | 80k shared between the 13 tasks | 80,000 |
qasc |
allenai/qasc | cc-by-4.0 | choice A..H | 9,060 |
refusal |
agentlans/refusal-classifier-data (english_only) | cc-by-4.0 | refusal noul, 30k | 30,000 |
snli |
stanfordnlp/snli | cc-by-sa-4.0 | as MNLI, 40k | 40,000 |
swarm_laya |
Aditharavind/swarm-laya-decisions | apache-2.0 | robot-swarm next action | 6,580 |
tasksource_jev |
tasksource/tasksource-jev-typed-decisions | other (per row, commercial-use licences only: Apache-2.0, MIT, CC-BY, CC-BY-SA, CC0, ODC-BY, BSD...; see the upstream license_use column) | commercial rows, 120k shared between sources | 119,595 |
tasksource_synth |
tasksource/synthetic-typed-decisions | apache-2.0 | aggregated soft answers | 25,186 |
tetris_laya |
aaung/laya-tetris-decisions | apache-2.0 | placement choice | 8,000 |
trading_laya |
Anay2010/laya-trading-decisions | apache-2.0 | synthetic trading states | 15,792 |
typed_synth |
n4ze3m/typed-decisions-synth | mit | teacher soft labels as targets | 25,859 |
when2call |
nvidia/When2Call | cc-by-4.0 | tool_call / request_for_info / cannot_answer / direct_answer | 24,776 |
Not republished:
typed_decisions_train: LocalLLaMA/typed-decisions is gated on the Hub.typed_decisions_bench: LocalLLaMA/typed-decisions is gated on the Hub (the benchmark test split).
Files
| file | rows |
|---|---|
data/allocine.parquet |
30,000 |
data/arc.parquet |
7,784 |
data/assistant_decisions.parquet |
15,455 |
data/banking77.parquet |
13,083 |
data/boolq.parquet |
12,697 |
data/calibrated_typed.parquet |
18,794 |
data/commonsense_qa.parquet |
10,962 |
data/hellaswag.parquet |
30,000 |
data/helmo_synth.parquet |
6,574 |
data/helpsteer2.parquet |
60,000 |
data/mnli.parquet |
60,000 |
data/neuro_laya.parquet |
16,456 |
data/notam.parquet |
9,051 |
data/open_jev.parquet |
47,800 |
data/pawsx_fr.parquet |
48,864 |
data/pngwn_typed.parquet |
51,019 |
data/procedural_typed.parquet |
80,000 |
data/qasc.parquet |
9,060 |
data/refusal.parquet |
30,000 |
data/snli.parquet |
40,000 |
data/swarm_laya.parquet |
6,580 |
data/tasksource_jev.parquet |
119,595 |
data/tasksource_synth.parquet |
25,186 |
data/tetris_laya.parquet |
8,000 |
data/trading_laya.parquet |
15,792 |
data/typed_synth.parquet |
25,859 |
data/when2call.parquet |
24,776 |
from datasets import load_dataset
ds = load_dataset("maxgfr/skycmd-decisions", "allocine")
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.