orcarouter/OrcaJev-Decisions

OrcaJev Typed Decisions — Full Corpus

One problem, one System-One decision: state + typed question → typed answer + calibrated probability, in a single forward pass, no decode loop.

Each case holds exactly one question, keyed q, of type choice, noul, or score. Gold is always verifiable or human/dataset-annotated — never model-generated.

At a glance

Cases	1,081,380  ·  train 1,069,380 / validation 6,000 / test 6,000
Domains	15
Question types	choice 66.0% · noul 19.4% · score 14.6%
State length	~77 tokens (p50), 942 (p90), up to 199,688
Choice options K	4 (p50), 39 (p90), up to 128
Positional shortcut	none — choice gold-position mean 0.4998

Question types

type	the decision	gold
choice	pick one of K listed options (2 ≤ K ≤ 128); a pointer over option spans	one option
noul	judge a single claim true/false ("the next step is to call X")	boolean
score	rate on a fixed ordinal scale (e.g. 1–5 stars, similarity buckets)	one level

Domains

Ordered by size.

domain	cases	what the decision is	gold
language	284,932	NLI / classification / QA / commonsense (GLUE, RACE, HellaSwag, WinoGrande, SciQ, DROP, ARC, …)	dataset label
games	225,517	best move / legal move / outcome on a text-rendered board (chess, gomoku, connect4, checkers, snake, 2048, tic-tac-toe)	verifiable oracle (engine, minimax, solver, rule)
agent	132,595	typed-decision supplements (tasksource, procedural, system-one, n4ze3m)	dataset gold
rating	108,434	ordinal score: review stars (Amazon/Yelp), response quality (UltraFeedback), semantic similarity (STS-B)	annotated rating
longhorizon	107,373	next step in a long multi-turn trajectory — which tool to call / reply, or next browser action — conditioned on the accumulated history (Nemotron interactive_agent, AgentTrek re-render)	the action actually taken
multilingual	54,100	NLI / classification / QA across languages	dataset label
retrieval	41,403	which passage answers the query, or needle-in-a-haystack over long context (MS MARCO, HotpotQA, QuALITY, BABILong, synthetic NIAH)	annotated relevant passage
preference	40,000	which of two responses is preferred (UltraFeedback)	preference label
knowledge	28,178	multiple-choice knowledge & truthfulness (MMLU, MMLU-Pro, TruthfulQA)	dataset answer
tools	27,624	which function/tool to call, or whether a call is needed (ToolACE, glaive)	annotated tool call
web	18,306	next action + target element on a web page (Mind2Web)	human-annotated next action
math	10,311	answer selection on math problems	verifiable answer
code	1,600	answer selection on code problems	verifiable answer
safety	624	refuse / harm-category of an agent request (AgentHarm)	human safety label
longcontext	383	answer a question over a long document (LongBench-v2)	dataset answer

Long-horizon = history-conditioned single-step decision, not rollout. Each case is one forward pass over the trajectory so far. AgentTrek is re-rendered with a compact past-action history plus only the current page's AXTree (kept within the token budget), so it lives in longhorizon, not web.

Quality pipeline

Applied to every case:

• normalize — dict states/instructions rendered to text; gold given an explicit type and a one-hot probability vector; noul forced to false/true.
• structural validate — 2 ≤ K ≤ 128, gold in the option set, non-empty state/instructions, probabilities normalized.
• length gate — drop states beyond ~200k tokens.
• content de-duplication — identity is state + option texts + gold text, so cases that differ only in option order or key labels collapse.
• choice option shuffle — option order randomized so the gold position is uniform, removing the positional shortcut (some sources list the gold first).

Choice gold-position mean 0.4998 (0.5 = no positional shortcut after shuffle) · noul true-share 0.463 · drops {"dup": 34}.

Statistics

By question type

type	cases
choice	713,998
noul	209,470
score	157,912

By source (67 sources)

source	cases
chess	149,997
rating	108,434
longhorizon	60,000
multilingual	54,100
procedural	50,000
tasksource_jev	49,675
agenttrek_lh	47,373
preference	40,000
knowledge	28,178
gomoku	22,000
n4ze3m	21,746
glaive	19,996
ms_marco	18,682
mind2web	18,306
snake	17,994
swag	15,000
qnli	15,000
fancyzhx/dbpedia_14	15,000
winogrande	15,000
hellaswag	15,000
hate_speech_offensive	15,000
toxic_conversations	14,989
student-question-categories	14,985
drop	14,867
enron_spam	14,301
ade_corpus_v2_classification	13,684
race	12,000
checkers	12,000
hotpotqa	12,000
connect4	11,971
systemone_dec	11,174
20news	10,993
sciq	10,481
math	10,311
legacy-datasets/banking77	10,003
tau/commonsense_qa	9,741
cola	8,543
allenai/qasc	8,131
subj	8,000
niah-synth	8,000
2048	7,953
toolace	7,628
sst2	6,911
wic	5,428
trec	5,381
amazon_counterfactual_en	4,998
allenai/openbookqa	4,916
mrpc	3,668
tictactoe	3,602
CR	3,389
allenai/ai2_arc	3,369
quality	2,523
rte	2,488
code	1,600
bbc-news	1,194
balanced_copa	1,000
wnli	635
agentharm	624
tweet_eval_stance_abortion	587
longcontext	383
cb	250
babilong	198

Format

• data/*.parquet (HF native, small row groups + page index) — id, domain, game, source, qtype, state, instructions, option_keys, options, label, probabilities, meta, plus case_json with the exact OrcaJev case. train is sharded (train-00000-of-00003.parquet, …, each ≤500MB); validation / test are single files.
• jsonl/*.jsonl — one OrcaJev case per line, ready for train_decoder (--mixture <dir with train.jsonl>).

This is a private repo — authenticate first (export HF_TOKEN=hf_... or hf auth login).

from datasets import load_dataset

# validation / test are small (6,000 cases each) — load fully into memory
val_ds  = load_dataset("orcarouter/OrcaJev-Decisions", split="validation")
test_ds = load_dataset("orcarouter/OrcaJev-Decisions", split="test")

# train is 1,069,380 cases / ~1.3GB — stream it instead of materializing it
train_ds = load_dataset("orcarouter/OrcaJev-Decisions", split="train", streaming=True)

import json
case = json.loads(next(iter(train_ds))["case_json"])

Loading a split with streaming=False (the default) still triggers download_and_prepare() for the whole default config the first time — i.e. requesting only split="test" will also download and cache train in the background. Use streaming=True for train to avoid that.

Provenance & license

Apache-2.0 for the generated data and pipeline code. Games are synthetic (chess positions from the Lichess puzzle DB, CC0). Web / tool / safety / agent / language / rating / preference / knowledge / retrieval / long-horizon / long-context cases derive from public datasets (Mind2Web, AgentTrek, ToolACE, glaive-function-calling, AgentHarm, tasksource, GLUE/SuperGLUE, RACE, HellaSwag, WinoGrande, UltraFeedback, MMLU / MMLU-Pro, TruthfulQA, MS MARCO, HotpotQA, QuALITY, BABILong, Nemotron-SFT-Agentic-v2, LongBench-v2, Amazon / Yelp reviews, STS-B, and others) — see each upstream dataset's own license and terms. Only text fields and the annotated decision are kept.