inclusionAI/LemonDecider-Decision-Mix

LemonDecider Decision Mix v1

1,199,692 independent typed decisions (choice / noul / score) mixed from 287 upstream datasets, with per-row upstream license evidence shipped alongside.

This is the training corpus of LemonDecider — it teaches a one-forward-pass decision engine: the model reads a state, is asked a typed question, and emits a probability distribution in a single readout — the chosen option for choice, a yes-probability for noul, a graded expectation for score. Every row is a single question with target supervision inherited from established upstream datasets.

The mix follows and extends the tasksource-jev-typed-decisions lineage: the same decision schema and recast philosophy, applied to a wider, curated mixture under explicit budget control.

Why use it

• Real supervision. Targets come from the upstream datasets' own labels, ratings, and annotator votes. Non-point-mass (graded) probability targets are preserved: 61,057 rows (5.09%) carry soft targets (choice 34,722 / noul 24,770 / score 1,565), from sources with mean ratings or ≥5-annotator vote shares. label is derived from target (first-index argmax on ties), not an independent gold annotation.
• Breadth with composition control. 287 upstream datasets across 864 task labels: commonsense and NLI, medical and science QA, sentiment, intent and routing, retrieval and grounding checks, safety and toxicity, aspect-based sentiment, procedural planning — over 40 dataset families and dozens of languages, under per-source caps so no source dominates.
• Clean by construction. Evaluation-suite protection fingerprints, cross-source deduplication, near-duplicate review, supervision-leak and positional-option checks, non-train split exclusion, commercial-license gating at every admission step, and full-file token-budget verification. Every cleaning step emits a count; nothing is dropped silently.
• License-honest. The public train.jsonl carries no provenance columns, but ships with row_licenses.jsonl: a per-row upstream identity and best-effort license_use classification aligned strictly by row order (verified by payload equality on all 1,199,692 rows). Commercial users can filter; everyone can audit.

Statistics

property	value
rows (train)	1,199,692
choice	1,012,354 (84.38%)
noul	182,326 (15.20%)
score	5,012 (0.42%)
choice options per row	2–235 (2: 313,688; 3: 217,158; 4: 215,729; …)
score scales	3 / 4 / 5 / 10 points
graded (non-point-mass) targets	61,057 (5.09%)
upstream datasets	287 (the tasksource collection expanded to 253 real upstream datasets, plus 40 direct roots admitted from official train/dev splits)
upstream license_use	commercial 793,739 (66.2%) / unspecified 372,275 (31.0%) / non-commercial 33,678 (2.8%)

Note that target being an array does not make it a soft label: [0, 1, 0] is a hard one-hot; only graded rows spread mass (e.g. [0.2, 0.7, 0.1]). Two-option rows can still be choice (not only yes/no).

Format

{"id": "row-0000001",
 "state": "Goal: Make Watalappan\nStep:",
 "question": {"type": "choice",
              "instructions": "Select the option that best answers the question.",
              "criteria": {"0": "Stir in 3 to 5 US tbsp (44 to 74 ml) of liquid if you want to thin the cereal.",
                           "1": "Squirt in equal parts of acrylic polymer gloss medium and white paint.",
                           "2": "Stir the hot water with the jaggery until it's smooth.",
                           "3": "Brush off the leftover wood shavings with a damp rag."}},
 "target": [0.0, 0.0, 1.0, 0.0],
 "label": "2"}

field	meaning
id	row-{1-based row number, 7 digits}; defines the public order
state	The text (and only text) to decide about; never reparsed or rewritten
question.type	choice, noul, or score
question.instructions	What to decide; the recast instruction of the source task
question.criteria	The option set, with 0-based keys preserving upstream order; absent when options carry no meaning (plain noul), explicit yes/no options for binary variants, scale points for score
target	Probability distribution over criteria (sum=1), scalar satisfaction probability for noul rows, or graded scale distribution for score
label	argmax of target, first index on ties (derived, not independent gold)

Rows are independent: the public file deliberately removes joint/group metadata. If you need multi-question states you must group by your own logic; none is recoverable from this file.

How it is built

Collection logic

The corpus is curated as one integrated mixture, sized to a one-day single-node training budget (≈1.2M rows at 8192 length), with per-source and per-suite caps throughout so that no source dominates. Its content comes together from three kinds of streams, merged and deduplicated as a whole:

• Decision-recast corpora. The bulk derives from the tasksource collection (488 configurations, expanded to the 253 real upstream HF datasets behind them — NLI and reasoning, QA and commonsense, sentiment, intent and routing, fact checking, safety, and dozens of languages), admitted under commercial-license and structural gates.
• Official benchmark training sets. For public leaderboards whose benchmarks ship genuine official train splits, those splits are included (e.g. MMLU auxiliary_train, HellaSwag, WinoGrande, ARC, ANLI, GSM8K, BANKING77, CLINC150, SGD, ContractNLI, HoVer, RAGTruth, ESCI, ACOS, iSarcasmEval, VAST, When2Call, MMLU-Pro, Humicroedit, BPoMP, cfcolor, Habermas Machine). Benchmarks with no official train/test division (BBH, POP909, CLadder, MuSR, SATA-Bench, New Yorker) are deliberately excluded — any subset of them would be test data; and where available, the splits are decontaminated against the evaluation sets by fingerprint. GSM8K distractors are rebuilt from other training problems' answers, so evaluation answers never leak as distractors.
• Method-level reproduction of a frontier mixture. The overall proportions reproduce, at method level, the published data-mixing approach of a frontier decision-engine training recipe — a replication of the composition methodology (and of the admission gates it described), not of its exact membership. It is an approximate reproduction by construction: no original row membership is claimed or recovered.

Acquisition practice. Every upstream snapshot is revision-pinned with size+SHA256 receipts and atomic publication; "file exists" is never acceptance — files must pass full-decode verification. Upstream identity (dataset, config, revision, license fields) is recorded per row in the internal ledger that powers the released license evidence.

Rewrite logic

A deterministic candidate adapter (13-field upstream typed schema → the public 5-field row) converts every source row alike:

• choice — criteria are the upstream option list re-keyed "0", "1", … in per-row order; instructions carries the task's own question when the inputs alone don't say what to predict, a generic instruction otherwise; target copies the upstream probability vector under a sum=1 check.
• noul — no options; target is the scalar yes-probability (entailment likelihoods, annotator agreement shares); label is derived at threshold p ≥ 0.5 and is explicitly marked as a derived threshold, not independent gold.
• score — an ordered scale kept in score semantics, with its distribution as the target (a mean rating of 3.4 on 1–5 puts 0.6 on 3 and 0.4 on 4).
• state — identity transform: the string is carried byte-identical, no reparsing, no truncation.
• Structural validation on every converted row — 1–255 options; no duplicate option strings; options that cite other options by letter or number keep their order; no supervision leak (gold answer / target / probabilities must not appear in the question text); native multi-question examples (e.g. 64 aspect-sentiment questions per review) are flattened into separate single rows.

Cleaning logic

Cleaning happens in layers, each with a machine-readable ledger:

1. Admission gates — non-commercial sources excluded at every admission step (both ours and the upstream collection's own filtered runs), non-train splits dropped, benchmark families blacklisted where their inclusion would contaminate evaluation.
2. Candidate review flags — positional options ("all of the above", "option A"), normalized-option collisions, duplicate ids, conflicting exact targets for identical inputs, image-bearing rows (text-only policy), and near-duplicates (5-word shingles; Jaccard ≥ 0.85, or containment ≥ 0.95 with ≥20 matches). Flagged candidates leave as review-marked records, not silent deletions.
3. Merge gates — cross-source exact-input deduplication first-seen-wins; evaluation-suite protection fingerprints (the public JevBench suite and the Decision-Index dev suites: sha256 of state+question, plus bare-state fingerprints with boilerplate states shared by many requests exempted); family isolation so one source's split family never straddles train/eval; token-budget gate with a whole-file tokenizer pass; then an equitable trim to the 1.2M budget, sampling proportionally by (source, suite).
4. Acceptance — every row is validated against a row contract at build time and the whole file re-validated at write time; the merged corpus has a frozen manifest and acceptance record (SHA-256 pinned).

The public single-record projection

train.jsonl is a strictly derived projection of the internal merged corpus: origin and joint-grouping metadata removed, ids reissued row-0000001…, only id/state/question/target/label kept. The projection was verified end-to-end: the source file's device, inode, size and mtime are asserted unchanged; every projected row was compared lockstep against its parent (payload and option-key order), 1,199,692 of 1,199,692 equal, zero mismatches; a brand-string self-scan over the entire file returns 0 matches; and validation.json records all of the above.

License and scope

This collection has mixed source terms, hence license: other. Original upstream dataset and annotation terms still apply. license_use is a best-effort classification: commercial, non-commercial or unspecified; it is not legal clearance. Missing or scoped evidence is retained explicitly.

The 1,199,692 rows draw on 287 upstream datasets. The public train.jsonl carries no provenance columns; row-level upstream identity and terms are aligned strictly by row order (verified by payload equality on every row: 1,199,692 / 1,199,692 vs. the internal parent mix). Each row resolves to exactly one upstream dataset, whose entry in the component-level evidence tables carries its license, use classification and link.

Row-level license_use distribution (see row_licenses.jsonl, upstream_aggregate.tsv and the full LICENSE-CARD section):

license_use	rows	share
commercial	793,739	66.2%
unspecified	372,275	31.0%
non-commercial	33,678	2.8%

To use only rows classified commercial:

import json
lic = [json.loads(l)["license_use"] for l in open("row_licenses.jsonl")]
ds = ds.add_column("license_use", lic)
ds = ds.filter(lambda use: use == "commercial", input_columns="license_use")

Where a row's embedded classification and its dataset-level table entry disagree, the stricter value is kept and the disagreement is recorded in the row's notes field, never silently reconciled (115,777 rows carry such retained evidence). Non-commercial components include facebook/anli (cc-by-nc-4.0), textmachinelab/quail, PKU-Alignment/PKU-SafeRLHF and BeaverTails, heegyu/webgpt_comparisons_ko, tasksource/sick, allenai/sciq, tasksource/contract-nli, jeggers/riddle_sense; commercial use must exclude these rows. This is a best-effort aid, not legal advice: check the original terms before relying on them.

Intended use and limitations

• Train typed decision engines (choice / noul / score heads) or any model that benefits from soft-target single-question supervision over diverse decision formats.
• Rows are independent single questions by design; group and joint-request structure is not recoverable from this file.
• label is derived from target; if you re-derive labels with a different tie policy you may disagree on ties.
• The mixture optimizes a one-day 4B-scale training budget, not coverage: proportions are budget artifacts as much as curation choices.
• Included official train splits come from public leaderboards' benchmarks; training on them is allowed practice but counters your held-out evaluation if you don't exclude accordingly.

Attribution

The upstreams are individually credited, with per-dataset links and terms, in upstream_aggregate.tsv.