Razvilka

Razvilka (развилка, "a fork in the road") is a Russian benchmark for structured decisions over text: given a text and a question with a closed set of options, pick the right one. It covers the decisions an LLM product makes all the time without generating text — routing, intent detection, topic and sentiment classification, moderation, relevance ranking.

• 735 items, 15 tasks, all in Russian, gold labels from published datasets.
• Four question types: choice (one of K options, 500 items), noul (yes / no, 139), score (a level on an ordinal scale, 46), ranking (order candidates, 50).
• Options are declared as text inside the request, so the same model is asked about a new label set without retraining.
• Every item went through review; disputed items were removed, nothing was relabelled.

Format

One row per item in data/test.jsonl:

field	meaning
id	item id
task	task name
type	choice, noul, score or ranking
state	the text the decision is about
request	the full request as JSON: {"state": ..., "questions": {<qid>: {"type", "instructions", "criteria"}}}
gold	JSON: an option key (choice, ranking — the best candidate), true/false (noul), or a level index (score)
source_dataset, source_config, source_split, source_row	where the text and label come from
license	licence of the source dataset

from datasets import load_dataset
import json

ds = load_dataset("artemsnegirev/razvilka", split="test")
item = ds[0]
request, gold = json.loads(item["request"]), json.loads(item["gold"])

Metric

Accuracy: the predicted option (top-1 for ranking, argmax level for score, p(true) > 0.5 for noul) must equal the gold. Argmax ties are broken lexicographically by option key. Chance level over the whole set is 0.257.

Evaluate

The scorer and an evaluation notebook live on GitHub: benchmarks/razvilka. Accuracy is computed overall, per question type and per task; argmax ties go to the lexicographically smallest option key.

Tasks

task	type	items	what is decided	source	licence
adverse_reaction	noul	50	does a post report an adverse drug reaction	cimm-kzn/RuDReC	not declared
domain18	choice	50	voice-assistant request → one of 18 domains	mteb/amazon_massive_scenario	CC-BY-4.0
emotion_choice	choice	50	emotion of a short text	mteb/CEDRClassification	Apache-2.0
entity_sentiment	choice	50	sentiment towards a named entity in a news sentence	dialogue-evaluation/RuSentNE-evaluation	MIT
grnti28	choice	50	scientific abstract → GRNTI rubric (28)	ai-forever/ru-scibench-grnti-classification	MIT
inappropriate	noul	41	is a chatbot utterance inappropriate (reputational risk)	mteb/InappropriatenessClassification	CC-BY-NC-SA-4.0
intensity4	score	46	emotion intensity on a 4-level scale	brighter-dataset/BRIGHTER-emotion-intensities	CC-BY-4.0
intents60	choice	50	user request → one of 60 intents	mteb/amazon_massive_intent	CC-BY-4.0
oecd29	choice	50	scientific abstract → OECD field (29)	ai-forever/ru-scibench-oecd-classification	MIT
relevance_list	ranking	50	rank 5 passages for a search query	miracl/miracl	Apache-2.0
sentiment3	choice	50	product review sentiment (3)	ai-forever/ru-reviews-classification	Apache-2.0
sentiment_film3	choice	50	film review sentiment (3)	ai-forever/kinopoisk-sentiment-classification	MIT
task_type35	choice	50	user instruction → task type (35)	ai-forever/POLLUX-instructions	not declared
topics6	choice	50	news headline → topic (6)	ai-forever/headline-classification	MIT
toxicity	noul	48	is a comment toxic	textdetox/multilingual_toxicity_dataset	OpenRAIL++

Licensing

Razvilka is a compilation: each item keeps the licence of the dataset it comes from, given in the license column and in the table above. The task wording and option descriptions are ours. "not declared" means the source publishes no licence. Items from inappropriate are CC BY-NC-SA 4.0 and may not be used commercially. Attribution for every source is in NOTICE.

Citation

TODO