lothnic/themis-decision-mixture-v1

Themis Decision Mixture v1

A curated text and image dataset for typed decision models. Each case provides evidence and explicit question criteria; targets are probability vectors aligned with saved option keys. The two configurations are alternative views of the same decisions, and must not be concatenated.

This release preserves every v0 decision, target, ID and split. New data emphasizes observed text weaknesses and adds finite-answer visual tasks. Counts below are measured exports, not collection targets. No training has been run with this release. Status: automatically validated candidate; independent human semantic review is pending.

The release contains 86,643 decisions, including 73,247 training decisions: 58,247 text-only and 15,000 image-grounded. The 16,830 image-grounded decisions across all splits use 10,072 unique images.

Actual composition

Split	Decisions	Image-grounded decisions	Packed requests
train	73,247	15,000	52,111
validation	4,336	491	3,148
calibration	3,883	602	2,801
test	5,177	737	3,943

The original v0 training set contributes 44,577 text decisions. New training additions:

Competency	Target	Retained
meeting summary matching	3,000	133
meeting decision action item detection	2,500	37
five level sentiment	3,000	0
contract entailment	2,500	2,500
credential placeholder disambiguation	2,000	0
product query relevance	2,000	2,000
general nli	3,000	3,000
routing agent control	6,000	6,000
document layout	4,000	4,000
diagrams technical	3,000	3,000
scene reasoning	4,000	4,000
charts tables	3,000	3,000
visual procedural	1,000	1,000

Zero or under-target rows are intentional shortfalls. Five-level sentiment and credential/placeholder additions are deferred because source rights or a defensible annotation mapping remain unresolved. No unrelated examples fill those quotas.

Data contract

state, parsed questions, and ordered images are the only model inputs. Text-only rows contain an empty image list. Images use Hugging Face List(Image) with original encoded bytes embedded in Parquet; no transient URLs or base64 strings are required.

questions, option_keys, metadata, and request-view targets are JSON strings to preserve dynamic question names and arbitrary option keys consistently in Arrow. Parse them with json.loads. In the decisions view, target is a numeric list. In the requests view, decision_ids align with question insertion order. metadata contains provenance and labels and must not be placed in model inputs.

Decision types:

• choice: nominal criteria keyed by option name.
• noul: yes/no, always [P(true), P(false)]. Source [no, yes] vectors are reordered.
• score: ordered levels, with target mass aligned to explicit indexed anchors; CLEVR counts use levels 0–10.

Hard source annotations and procedural ground truth remain one-hot. Inherited soft distributions are preserved. No uniform distributions are substituted for missing labels.

import json
from datasets import load_dataset
rows = load_dataset("lothnic/themis-decision-mixture-v1", "decisions", split="train", streaming=True)
row = next(iter(rows))
questions = json.loads(row["questions"])
keys = json.loads(row["option_keys"])
assert abs(sum(row["target"]) - 1) < 1e-5
inputs = {"state": row["state"], "questions": questions, "images": row["images"]}

Sources and conversions

• Inherited Themis v0: revision 92843bdf66de205d6b4238c6a0d70701fd6737a7; original mixed licenses and limitations remain applicable.
• ContractNLI: complete contracts and native entailment/contradiction/not-mentioned annotations. Overlong decisions are excluded with their evidence intact; evidence spans are not used to crop contracts.
• Amazon ESCI: native US small-version query/product judgments, joined by product ID and locale. Native held-out product/query identities and v0 query/title matches are excluded from new candidate sampling.
• AMI 1.6.2 manual annotations: summary matching derives from native extractive-to-abstractive links. Other same-meeting summaries are distractors, not independently proven false entailments. Action/decision classification uses summary sentences and native summary sections, not a validated utterance-level action/decision/neither task. Unlinked utterances are never labeled neither.
• WANLI: native three-way NLI annotations grouped by seed MNLI pair ID and connected normalized premises. Native WANLI test seed IDs and premises are excluded. Its generated-and-human-revised annotation process must not be described as purely human authored.
• Open-Jev mailroom-control-v1: original synthetic routing and evidence-control cases; original family groups and native boundaries survive conversion. Counterfactual labels are procedural, not human probabilities.
• AI2D: native diagram questions and answer choices; all questions from an image stay together. CC BY-SA 4.0 obligations continue to apply.
• DocLayNet region: native highlighted-region classification over eleven classes. The bounded selection in this release contains only financial-report pages, not all DocLayNet document domains. Different highlight overlays connect through original page hashes and document filenames. No page-absence labels are inferred from sampled regions.
• CLEVR: native scene QAs, converted only through complete known finite answer vocabularies; max four selected questions per scene. Original functional programs are not supplied by this wrapper.
• Paper folding: fresh procedurally generated native five-panel choices; complete diagrams retained. Generator original assets are Apache-2.0 under the source evidence; upstream shipped benchmark data separately uses CC BY 4.0, and its notices are retained. This is not an ImageNet/photo subset.
• Themis charts v1: original seeded bar, horizontal-bar and line charts with exact count labels. Tables, seeds and labels are provenance only. Style 3 is entirely test-only. This small grammar does not establish broad chart/table/map competence.

Exact revisions, input checksums and acquisition routes are in sources.lock.json. Source notices and attribution are in ATTRIBUTION.md, LICENSE.md and attribution/.

Splits, overlap and limits

Native held-out boundaries are preserved. Additional train-derived groups are assigned deterministically with seed 3407 (90% train, 3% validation, 3% calibration, 4% test), before quotas. Images, original document parents, scene identities, meeting series and source counterfactual families connect related decisions. Components spanning different proposed splits are quarantined. Selected decisions are packed without truncation; every decision appears exactly once per view.

Protected benchmark material includes local Decision Bench, SysOne and Typed-Decisions test records, plus v0 holdouts. Checks use known source-parent IDs, normalized exact states/leaves, and literal 32-word fragments. A match quarantines the new connected group. Conservative literal filtering can remove legitimate boilerplate. These checks do not prove semantic deduplication, independent benchmark validity, or absence from base-model pretraining. Existing v0 is inherited unchanged, including its previously reported overlap limits. Benchmark results used to choose priorities are development-informed.

The 2B preparation profile uses the pinned tokenizer/image processor, a 4,096-token combined budget and at most 1,024 visual tokens/image. Original assets remain intact; training preprocessing may resize them proportionally. Decisions over budget are rejected rather than truncated. Dense text and small visual details may remain difficult after preprocessing; semantic and visual-reliance reviews remain pending.

Validation and intended use

audit/ records image decoding/checksums, new-component split overlap, probability alignment, token exclusions, packing coverage, and all-field Parquet round trips. These are automated data checks, not evidence of improved model accuracy. Use only the train split for optimization, validation for model selection, calibration for fitting calibration parameters, and test for final reporting.

All inherited tasks retain their original caveats. Source-derived and synthetic labels can be ambiguous or wrong. This dataset does not certify decisions for consequential deployment. No model performance is claimed by this dataset release.

Reproduction

See BUILD.md, manifest.json and the copied builder under builder/. The recorded bounded acquisitions avoid downloading the full combined vision pool. Individual licenses remain in force; there is no blanket relicensing of third-party records.

Measured coverage limits: new AMI training rows span only three meeting series; this is a small development addition, not broad meeting-domain coverage. See audit/coverage.json for group and target counts.

The document source uses expanded export a023fe646d5ba2fc83fcc553f356f9bba58d06b1, training shards 0–3, superseding the earlier research note about the old 1,000-row export. sources.lock.json is authoritative for this actual build.