yyhlm/typed-decisions-ru

Typed Decisions (Russian)

Русская версия бенчмарка LocalLLaMA/typed-decisions: модель получает одно состояние (state) и отвечает сразу на несколько типизированных вопросов о нём, причём каждый ответ — распределение вероятностей, а не одна метка. Тексты переведены на русский, разметка (gold) — та же, что в оригинале.

A Russian translation of LocalLLaMA/typed-decisions. Every case, split and gold distribution is the original's; only the natural-language text is translated. For what the benchmark measures, how it was built and how to read a score, see the original card — all of it applies here.

What was translated, and what was not

Translated from revision f7a2487 of the original, by machine translation with Claude Sonnet 5.5 (Anthropic).

Column	Status
state	translated (JSON structure and keys kept; string values in Russian)
questions	instructions and option descriptions in criteria translated; question ids and option keys (continue, human_review, …) kept
factors	text values translated (e.g. task and constraint descriptions); not model input
id, workflow, split, n_questions	unchanged
gold	unchanged, byte for byte
label_agreement	unchanged, byte for byte

Checked programmatically against the original: all 1,600 case ids match, gold and label_agreement are identical for every case, every question keeps its id, and every state was translated.

Option keys stay in English on purpose: gold is keyed by them, so a row still scores against the original gold without any mapping, and state + questions remain exactly the body of a POST /v1/systemone request.

Splits

Workflow	Train	Test
agent_trace_observability	300	100
customer_service	300	100
invoice_processing	300	100
security_incidents	300	100
all	1,200	400

Each case asks 5 questions over one shared state: 2,000 decisions in test. test is the benchmark.

Usage

from datasets import load_dataset
import json

ds = load_dataset("yyhlm/typed-decisions-ru", "all", split="test")
row = ds[0]

state = json.loads(row["state"])          # Russian text
questions = json.loads(row["questions"])  # Russian instructions and option descriptions
gold = json.loads(row["gold"])            # unchanged from the original

print(questions["action"]["instructions"])
print(gold["action"]["probabilities"])

Caveats

• The gold was produced on the English text. It is the mean of three teacher samples on the original cases, not a relabelling of the Russian ones. Where the translation shifts meaning, the gold does not follow. Treat a score here as agreement with the English teacher on translated input.
• The translation is machine-made (Claude Sonnet 5.5). Expect occasional awkward phrasing or terminology that a human translator would choose differently.
• Scores are not directly comparable with the English leaderboard. Same cases and gold, different input language.
• Say which mode you evaluate in. As in the original: a specialist is trained on this dataset's train split; a generalist has never seen these workflows. The two are not comparable.
• The data is synthetic, as in the original.

License and attribution

Apache License 2.0, the license of the original dataset (see LICENSE).

Based on LocalLLaMA/typed-decisions (Apache-2.0). Changes: the text fields listed above were translated from English into Russian; nothing else was modified. This translation is not affiliated with the authors of the original dataset.