lothnic/themis-decision-mixture

Themis Decision Mixture · v0

Training data for models that answer typed questions with probabilities over the allowed answers.

Themis Decision Mixture v0 contains 54,046 text-only decision instances curated from annotated tasks, procedural problems, teacher-labeled workflows, and real code changes. Each instance pairs a case with an instruction-defined question, answer options, and an aligned target distribution. The dataset is curated for the planned Themis text-only decision model based on Qwen3.5-0.8B, and can also support other schema-conditioned decision models.

This is a data release. It does not include trained model weights or claim a model accuracy result.

Training decisions	Soft targets in training	Decision types	Covered categories
44,577	12,900	choice · noul · score	10

Splits and two views

Split	Decision instances	Packed requests	Requests with multiple questions
train	44,577	33,573	5,324
validation	3,170	2,275	343
calibration	2,606	1,945	218
test	3,693	2,781	420
Total	54,046	40,574	6,305

• decisions is the default configuration: one question per row, useful for sampling, inspection, and individual-decision training.
• requests groups questions about exactly the same state into bounded multi-question requests. Every decision appears exactly once in its split's request view.

These configurations are two representations of the same labels. Concatenating them would duplicate training supervision. Keep test untouched; use validation for model selection and the separate calibration split for fitting probability calibration.

Load and reconstruct a request

import json
from datasets import load_dataset

train = load_dataset("lothnic/themis-decision-mixture", "decisions", split="train")
row = train[0]

model_input = {
    "state": row["state"],
    "questions": json.loads(row["questions"]),
}
target = row["target"]
option_keys = row["option_keys"]
provenance = json.loads(row["metadata"])

assert len(target) == len(option_keys)
assert abs(sum(target) - 1.0) < 1e-6

For multi-question requests:

requests = load_dataset("lothnic/themis-decision-mixture", "requests", split="train")
row = requests[0]

model_input = {
    "state": row["state"],
    "questions": json.loads(row["questions"]),
}
targets = json.loads(row["targets"])
option_keys = json.loads(row["option_keys"])
decision_ids = json.loads(row["decision_ids"])

Only state and decoded questions are model inputs. Targets, category tags, provenance, and sampling metadata must remain outside the input. These are probability-supervision files, not ordinary chat SFT completion files; a trainer must consume the target distributions explicitly.

Decision types and target alignment

Type	Meaning	Target convention
choice	Pick one supplied option	Distribution in option_keys order
noul	Answer a yes/no question	[P(true), P(false)], with keys ["true", "false"]
score	Judge an ordered rubric	Distribution over level indices "0", "1", … in the original scale order

The training split contains 30,400 choice, 7,292 yes/no, and 6,885 score decisions. Soft target distributions are preserved. Only rounding drift within 1e-5 of a unit sum is normalized; malformed distributions are rejected.

Dataset columns

Variable question names and answer-option keys are stored as JSON strings so Parquet and the Dataset Viewer have a stable schema. Decode those columns with json.loads; do not treat the encoded JSON as an extra instruction.

decisions column	Contents
id	Stable decision identifier
state	Case text; structured cases may themselves be JSON text
questions	JSON object containing the typed question, instructions, and criteria
target	List of target probabilities
option_keys	Ordered keys aligned with target
category, kind	Coverage category and decision type
source, source_dataset	Underlying task name and immediate source collection
label_origin, license	Label provenance and recorded source terms
payload_tokens, split	Tokenized request size and assigned split
metadata	JSON provenance: pinned revision, source identity, grouping, fingerprints, and transformation details

The requests view contains id, state, questions, targets, option_keys, decision_ids, n_questions, payload_tokens, split, and metadata. Target vectors and original decision IDs are keyed by question ID. The byte-identical audited JSONL exports are also available under original/.

Training coverage

The mixture follows practical decision-use categories. Counts are actual selected decisions, not a claim of equal capability across domains.

Category	Training decisions
Engineering	7,000
AI agents and evaluations	7,000
Trust and safety	4,000
Customer support	5,000
Sales and commerce	3,595
Finance	3,500
Legal	5,000
Product management / feedback triage	482
Data and analytics	5,000
Documents and meetings	4,000

The original target was 50,000 training decisions. 5,423 places remain unfilled: commerce 405, finance 1,500, product management 2,518, and text-based design 1,000. Missing categories were not padded with unrelated tasks. No design examples are included in v0.

Sources and label provenance

The immediate training sources are:

• Tasksource Jev Typed Decisions: curated source annotations and deterministic task transformations.
• Procedural Typed Decisions: multi-question problems with rule-derived targets.
• Synthetic Typed Decisions: workflow judgments with teacher distributions.
• LocalLLaMA Typed Decisions: service, invoice, security, and agent-observability workflows.
• CommitPackFT: Python code diffs paired with their original commit messages.

Training label origin	Decisions
Source annotation or deterministic transformation	20,285
Teacher distribution	12,224
Procedural rule	6,215
Original code-diff / commit-message pair	5,853

Full source revisions and shard checksums are in audit/sources.lock.json. Original task identities, source dataset references, and recorded terms are retained in attribution/source-catalog.json and each row's metadata. See ATTRIBUTION.md for source credit.

Curation and validation

Curation uses deterministic sampling with seed 3407, category budgets, source caps, and rotation across question types and labels. Redundant augmentation variants are removed. Procedural records embedded in the broad Tasksource mixture are skipped in favor of the native procedural collection. Duplicate choice permutations are collapsed with probability alignment preserved; conflicting duplicate decisions are excluded.

Native held-out records are kept out of training. Connected case/state groups stay in one release split. Synthetic examples retain workflow grouping. CommitPackFT fork aliases are connected before assigning project-level splits; three distractor messages are drawn from the same derived partition, and option provenance is retained.

Decision Bench is used solely as an overlap-exclusion reference. 28 literal overlap matches were removed. Its benchmark rows are not included as Themis training or test examples.

The completed audit checked all 54,046 decisions and their packed request counterparts for:

• Valid probabilities and exact alignment with the pinned original source targets.
• Separation of input fields from labels and provenance.
• No state, connected-group, or decision-fingerprint overlap across release splits.
• No exact native held-out states in training.
• The configured literal Decision Bench overlap checks.
• Preservation of every decision and target during multi-question packing.
• Payload token budgets and output-file checksums.

The curation code's 22 regression tests passed at release. Audit results and the human-readable report are included under audit/.

Requests are bounded to 1,920 tokens of canonical API-request JSON using the pinned Qwen3.5-0.8B tokenizer, with 128 additional tokens reserved against a 2,048-token starting budget. Evidence is not truncated to meet that budget. A trainer must still re-check its actual prompt template and head-specific markers.

Limitations

• This is an automatically curated pilot, not an independently adjudicated gold standard. Structural validation does not prove semantic correctness. Teacher distributions reflect their teachers' judgments, and procedural targets were not independently recomputed from the generators.
• Holdouts are not universally unseen schemas. 270 validation, 277 calibration, and 317 test exact question/criteria schemas also appear in training. Workflow grouping and new cases should not be presented as a universal zero-shot schema evaluation.
• Exact states and normalized literal 32-word fragments are checked against Decision Bench. These checks do not rule out paraphrase overlap, all cross-dataset overlap, or exposure during backbone pretraining.
• Category mappings use explicit source rules and synthetic application/skill heuristics. Product management mostly covers feedback triage; documents coverage includes entailment and dialogue acts. These are not comprehensive product-planning or meeting-action extraction benchmarks.
• Code-message distractors may be easier than semantically close alternatives. Engineering coverage does not establish competence on every engineering decision type.
• The release is text-only and is not language-filtered; multilingual source examples may be present. It does not reproduce image-dependent design, receipt, or chart tasks.
• Some source material contains offensive language or adversarial instructions, especially in safety tasks. Such text is dataset evidence, not an instruction to the dataset consumer.

License and attribution

Mixed source terms apply. There is no blanket Apache-2.0, MIT, or CC-BY license for this collection. Recorded terms include attribution, share-alike, code, and database licenses; GPL-tagged material is also present. Exact source strings, including upstream ambiguities, are preserved rather than silently replaced.

The curation filter excluded noncommercial, unspecified, bare other, and custom-scoped terms, using upstream metadata as evidence. That filter does not independently establish rights for every source component. Consult LICENSE.md, ATTRIBUTION.md, and the per-row provenance for the applicable terms and original source references.

Citation

@misc{themis_decision_mixture_v0,
  author = {{lothnic}},
  title = {Themis Decision Mixture, v0},
  year = {2026},
  url = {https://huggingface.co/datasets/lothnic/themis-decision-mixture}
}

Please also credit the original datasets behind the records you use. Themis is a curation of their work, with provenance retained for that purpose.