mshojaei77/tasmim-training-mixture-v1

Tasmim Persian-English Typed-Decision Training Mixture v1

A multi-source training corpus converted into Tasmim's typed-decision case schema. Each case carries a shared inference state, one or more aligned questions, type-correct gold targets, and source-level provenance. The release combines Choice, Noul, and ordinal Score decisions from Persian and English sources.

Size and splits

Split	Cases	Choice decisions	Noul decisions	Score decisions	Total decisions
train	275,026	210,341	84,044	28,658	323,043
validation	13,480	10,633	3,735	1,270	15,638
calibration	9,860	7,742	2,801	960	11,503
test	9,862	7,742	2,802	949	11,493

Total: 308,228 unique exact-state cases and 361,677 decisions across 23 tasks. Choice: 236,458 (65.4%); Noul: 93,382 (25.8%); ordinal Score: 31,837 (8.8%).

Language tags are content-level case tags: 232,134 Persian cases and 76,094 English cases. Their question counts are 283,663 Persian and 78,014 English decisions. A multi-question case is counted once as a case and once per question as a decision.

Configurations

• all is the default and contains every retained license band.
• permissive contains only the permissive band.
• sharealike isolates share-alike rows.
• noncommercial_sharealike isolates CC BY-NC-SA rows.

The all config has mixed source terms; see license and provenance details before using it. The license: other metadata reflects that no single license can accurately cover all included rows. Source-specific attribution, ShareAlike, and non-commercial obligations remain attached to their rows.

Load the dataset

from datasets import load_dataset

ds = load_dataset("mshojaei77/tasmim-training-mixture-v1")  # defaults to all
permissive = load_dataset("mshojaei77/tasmim-training-mixture-v1", name="permissive")
sharealike = load_dataset("mshojaei77/tasmim-training-mixture-v1", name="sharealike")
noncommercial = load_dataset("mshojaei77/tasmim-training-mixture-v1", name="noncommercial_sharealike")
train = ds["train"]

Schema and provenance

• state: JSON string of inference-time fields.
• questions: JSON map of question IDs to Choice, Noul, or ordered Score specifications.
• gold: JSON map aligned to question IDs; each target includes a complete one-hot probability distribution.
• task_ids_by_question: task ID aligned to each question.
• provenance_json: source revisions, source row references, license terms, label origin, and per-question provenance for coalesced cases.
• language, supervision_tier, license_band, source_id, and source_revision support filtering and audits.

See the 88-entry source audit, task and label coverage, machine-readable registry, full validation results, and reproduction instructions.

Quality checks

• Train-only task-majority test baseline: 0.433 over 11,493 decisions.
• Train-only question-only test baseline: 0.453; exact question-feature coverage 90.6%.
• Exact normalized input-signature collisions with pinned Tasmim-Bench v1.1.0: 0.
• Lexical near-duplicate candidates at 92 token-set ratio: 0; the screen requires a shared rare five-token shingle and cannot detect all semantic paraphrases.
• Retained direct contact-pattern matches in state fields: 0.
• Semantic family and exact normalized state split-crossing counts are zero.
• All gold vectors are complete one-hot targets; every Score level is present; every Noul task has both true and false targets.

Shortcut review

• parsinlu_aspect_sentiment.aspect: Cramer's V 0.2948 (n=13,609); retained as a legitimate input field, not removed based on association alone.

Limitations and training readiness

• No labels are human-verified by this project. The release includes source labels, LLM-assisted/automatically labeled data, and explicitly synthetic deterministic toy-policy outputs; tiers are separate and preserved.
• No teacher-only rows are included. Synthetic oracle decisions must not be interpreted as real-world or human-preference truth.
• Only 23 decision tasks are represented; most of the 88 catalog entries were excluded because of evaluation-only status, missing/unclear license or labels, restricted terms, unavailable data, or unsupported schema.
• The JEV tool-use sources have dynamic option sets; 2,804 option names appeared as candidates but were not observed as targets in this release. Per-row targets remain aligned to that row's candidate set; see validation results.
• Source 38's binary codes remain opaque as 0 and 1; this release does not infer their polarity meaning.
• Persian source diversity is uneven. Some corpora are short reviews, intent requests, scientific abstracts, synthetic policy examples, or generated behavior scenarios; they are not a representative sample of Persian speakers or all domains.
• The public Tasmim-Bench was used only as a pinned text/signature overlap screen. No benchmark rows, labels, or predictions were included in training. This is not a claim of generalization to an untouched evaluation set.
• Use the provided splits for dataset experiments, but do not assume the public-source components are independent or human-verified.

Reproducibility

Training-source release reference: mshojaei77/tasmim-data-v1 at 9b8ca960fd2c1c92b15c8ec047302265c4c2adee. Benchmark overlap reference: mshojaei77/tasmim-bench at c7b48b97dcbd2bfd243551ec9482cb314fae27fd. The local builder uses pinned source revisions listed in docs/SOURCE_REGISTRY_88.json.

This repository contains data and preprocessing reports only. No model was fine-tuned as part of this release.