solar-decisions-corpus-v4

816,741 fully-labeled System One / typed-decision examples (choice · score · noul), labeled end-to-end by upstage/solar-decide-flash via the OpenRouter Decisions API. Every target is a full calibrated probability distribution — not hard labels.

License: Apache-2.0 (matches corpus v3) Teacher: upstage/solar-decide-flash ($0.05/M input, free output tokens) Total labeling cost: $20.65 across 816,741 API calls · 0 errors Labeling window: Oct 5–6, 2026 · 42.4 rows/s sustained over 5h21m

---

What is this?

Corpus v4 is a full relabel of the Jev-distill-corpus-v3 dataset with a different teacher, plus 120,000 brand-new questions across 24 fresh domains that v3 never touched. The result is a second opinion on every row: same states, same questions — fresh, independent calibration.

v3 was labeled by Jev 1.13 (TypeSafe). v4 relabels the exact same 696,741 rows with Solar Decide Flash and adds a large slice of new territory. Researchers can use v3 + v4 as a paired multi-teacher corpus — identical inputs, two independent probability sets — to study teacher disagreement, calibration transfer, and cross-teacher distillation.

The v3 banks, relabeled (696,741 rows)

Same states and questions as SargeDev/jev-distill-corpus-v3, keeping v3's original train/validation/calibration/ood split assignments. Stream sources (unchanged from v3):

Stream	Rows	Origin
yuri_v3 synthetic	462,920	generated synthetic decision states, Jev-labeled in v3, Solar-relabeled here
yuri_v1	142,935	the clean v1 corpus (synthetic memory-relevance and domain states)
Open-Jev (ZefanCai)	94,801	CC0 redistribution via v3, Solar-relabeled
workflow-controls / game states / other structured	rest	CC0 Open-Jev companion states

Fresh v4 domains (120,000 rows)

24 procedurally generated domains with zero name/event/subject overlap with v3's 53 domains (verified programmatically). New families and domains:

• transport (15,000): aviation ground ops, maritime ops, fleet dispatch
• industrial (10,000): trades diagnosis, welding & NDE
• consumer (25,000): retail ops, vehicle service, hospitality, pest control, petcare
• financial (10,000): insurance claims, credit underwriting
• public (10,000): municipal permits, emergency management
• agri_food (10,000): agricultural field ops, food safety
• games (15,000): tabletop adjudication, content moderation, esports meta
• arts (10,000): performing arts, sports officiating
• education (10,000): education operations, personal finance ops

Each v4 state fuses 2 events per state with context suffixes (same renderer conventions as v3's fixed renderer), giving ~1,667 distinct combo templates per domain. Deduped by sha256 of (state + question).

---

Schema (drop-in compatible with v3)

{
  "id": "ours_8ed01ce07cd55547 | v4_...",
  "domain": "maritime_ops",
  "family": "transport",
  "kind": "choice | score | noul",
  "question": "Which action should the operator take first?",
  "options": ["A", "B", "C", "D"],
  "state": "The vessel is 3 nm from harbor ...",
  "target": [0.32, 0.41, 0.15, 0.12],
  "teacher_confidence": 0.72,
  "source_split": "v3_relabel | v4_fresh",
  "v3_split": "train (v3 rows) | ''",
  "has_v3_jev_label": true,
  "split": "train | validation | calibration | ood | test"
}

Kind semantics (identical to v3):

• choice — pick best of N options; target = probability per option (order matches options)
• score — 0–5 rubric judgment; target = probability over labels "0".."5"
• noul — yes/no binary; options ["false","true"]; target = [p(false), p(true)]

All targets renormalized to sum 1.0.

---

Files

File	Rows	Size	SHA-256
train.jsonl	756,619	585.4 MB	c8efd6a13dd81e372b1b841682c4fdabc45f54ef99063cd1dc2bd44ff613042c
validation.jsonl	18,813	14.4 MB	2b3adc78738bec7cf02e7728a3103408fb2614ba69c6eee820b49028a944c987
calibration.jsonl	13,589	10.9 MB	c0f35ec02ba3f56f3988b46a1835c48bee8611eb6eee7fb0d36cb49a5c5f3361
ood.jsonl	17,720	18.8 MB	8987dc21ee973704885cb5806c28b2be4053d34b4b363e0e02a64b86923a995a
test.jsonl	10,000	7.8 MB	850d9e3ccd1d6151f50ab7d2fbae34092bffc6df68078f6c64d26ccd0b93d373

Splits: v3 bank rows keep their original v3 split (train / validation / calibration / ood). v4 fresh rows are 92% train / 4% validation / 4% ood. test.jsonl is a fresh frozen 10k test set, stratified proportionally across every (kind × source) group — disjoint from all training material.

Kind distribution

Kind	Rows
noul (yes/no)	406,165
choice	213,900
score	196,676
Total	816,741

Families (90 domains total)

Family	Rows		Family	Rows
knowledge	206,512		theology	24,174
openjev	90,886		biology	24,150
agent	60,271		spatial	23,962
technical	60,199		chemistry	15,466
business	49,794		transport	15,000
medical	40,265		games	15,000
genomics	32,671		agri_food	10,000
consumer	25,000		financial	10,000
structured	24,593		education	10,000
science	24,550		public	10,000
physics	24,248		arts	10,000
			industrial	10,000

---

Quality & Calibration Findings

Full 816,741 rows completed with 0 API errors.

Solar-vs-Jev agreement (full v3 bank vs v3 gold)

Kind	Metric	Value
noul	MAE, Solar p vs Jev p(true)	0.304 (45% within ±0.25)
score	MAE, Solar score vs Jev expected value	0.491 (57% within ±0.5)
choice	same-argmax rate	58% (n=173,886)

Stable from n=100 (S0 probe) through n=816k — no drift. Interpretation: Solar Decide Flash is a genuinely distinct, internally consistent teacher — not a Jev echo, not noise. The 42% choice-argument disagreement and the moderate continuous-space deltas (noul 0.30, score 0.49) are real calibration differences between two independently trained decision models, which makes this pair valuable for:

• multi-teacher distillation (soft targets from two schools of calibration)
• teacher-disagreement research (paired rows share identical inputs)
• calibration-transfer studies

Calibration shape (63k-row sample)

Kind	mean max-prob	%>0.9	%>0.99
noul	0.846	49%	1.2%
choice	0.643	11%	0.5%
score	0.486	0%	0%

Soft, well-distributed probability mass — no one-hot collapse, no saturation. Ideal for knowledge distillation (entropy stays in the targets).

---

Data quality / privacy

• 0 real secret hits across the full corpus on a pass built for real token shapes (sk-…48+, hf_…34+, AWS AKIA, GitHub ghp_, Slack xox). (A first-pass grep flagged 5 rows, all URL-slug substrings like sk-the…card — false positives, confirmed by context inspection.)
• 0 rows carry Will-personal data. 5 rows contain the name "Karina" — all verified to be public web content (a 1985 film's credits, a public medical Q&A about NGT feeding, a doctor's public profile link, a Family Guy episode description, a Swedish rock band bio). Per the project's redaction-scope rule, public-content name collisions are preserved for generalization; nothing traces back to Will personally.
• 34 distinct email addresses appear — all public/example addresses in public QA or wiki-style text (support@disneyworld.com, info@ittanta.com, etc.). Per redaction-scope rule: example/public emails are retained intentionally.
• Deduplication: sha256 (state.lower + "||" + question.lower) → 0 duplicates within this corpus. Dedup against v3 was structural (the corpus IS v3's row set, relabeled, so 1:1 state matching with v3 holds by design).

How to use

Works with the same loaders as SargeDev/jev-distill-corpus-v3:

from datasets import load_dataset
ds = load_dataset("SargeDev/solar-decisions-corpus-v4")
# or paired research (identical ids → join v3 and v4 targets on `id`)

Attribution

This dataset was created and maintained by SargeDev and contributors for open research into learned System One / calibrated decision-making. If you use this dataset for training, fine-tuning, evaluation, or derivative models, please credit SargeDev/jev-distill-corpus-v3 and SargeDev/solar-decisions-corpus-v4, and link back to this dataset.

Credits

• Teacher: Solar Decide Flash via OpenRouter Decisions API
• Prior work: v3 (Jev 1.13 labels); Open-Jev (ZefanCai, CC0)
• Method: same 32-worker sharded labeling architecture as v3 (gen/solar_label_v4.py) · stratified 34-choice/33-score/33-noul 100-row S0 probe validated schema before spend · hard cost ceiling $22 enforced (final: $20.65)