datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
moe-routing-drift-results
MoE routing drift — results
Measurements for a 2x2 experiment: adaptation (none / GEPA / prompt-tuning /
prefix-tuning) crossed with router retraining (frozen gate / gate retrained), on
inclusionAI/Ling-mini-2.0 and Qwen/Qwen3-30B-A3B-Instruct-2507. The weights are in
moe-routing-drift-checkpoints.
Content warning. quality/*/*.responses.jsonl contain verbatim comments from
civil_comments together with model outputs; the task is toxicity labelling, so the text
includes insults… See the full description on the dataset page: https://huggingface.co/datasets/AverageMetaheuristicsEnjoyer/moe-routing-drift-results.SPROUTscugnizz-v22-routing
scugnizz-v22-routing
Synthetic routing data including correct no-tool behavior and deterministic tool selection.
Format: Hermes/OpenAI-style messages plus tools.
RoutingCompendium-perf
RoutingCompendium — Performance
Five public LLM-routing benchmarks, from 10 to 112 candidates each, harmonized into one schema with prompt embeddings included. Companion to Wikit/RoutingCompendium-cost, which holds the inference price of each candidate.
Built for "Targeting the Exceptions: Online Active Learning for Retrieval-Based LLM Routers". Code: github.com/wikit-ai/online-active-learning-llm-routers
Why this dataset
Every benchmark below ships its own… See the full description on the dataset page: https://huggingface.co/datasets/Wikit/RoutingCompendium-perf.Prompt-Routing-DatasetPrompt Routing Dataset · Multi-Task Infrastructure Routing
About this dataset
This dataset is a highly dense, premium alignment asset explicitly designed to train Edge Orchestrators and Routing Models ranging from 50M to 1.5B parameters.
When deploying small language models (SLMs) on consumer hardware or local edge instances, running multi-step mathematical derivations or complex architectural software tasks often causes catastrophic hallucinations or syntax breakdown. This dataset… See the full description on the dataset page: https://huggingface.co/datasets/SupraLabs/Prompt-Routing-Dataset.commitmoe-qwen35-fp8-expert-routing-tracesarabic-prompt-routing
Arabic Prompt Routing — توجيه عربي صفري
233,720 rows. Each row is a text, a set of free-text categories, and which one it belongs
to. Categories are arbitrary Arabic — the point is a model that routes into a label set it has
never seen.
بالعربية: مجموعة بيانات عربية لتوجيه النصوص إلى فئات يكتبها المستخدم بلغة طبيعية.
الفئات ليست ثابتة، والهدف نموذج يوجّه إلى فئات لم يرها أثناء التدريب.
Arabic counterpart to the task in
LiquidAI/LFM2.5-Encoder-350M-Prompt-Router.
split… See the full description on the dataset page: https://huggingface.co/datasets/oddadmix/arabic-prompt-routing.RoutingCompendium-cost
RoutingCompendium — Cost
Inference price of every candidate LLM appearing in Wikit/RoutingCompendium-perf.
The two datasets are meant to be loaded together: -perf gives what each candidate scores on a query, -cost gives what calling it costs.
Splits
One split per benchmark, with the same names as RoutingCompendium-perf (RouterBench, Sprout, EmbedLLM, FusionBench, R2Bench). Each split lists the candidates of that benchmark's pool — a few dozen rows at most.… See the full description on the dataset page: https://huggingface.co/datasets/Wikit/RoutingCompendium-cost.SPROUT-o3minisemantic-repair-routing
semantic-repair-routing
The 84,819 supervised pairs that trained
SemanticRepair-270M:
a message somebody actually wrote, and the requests inside it restated
plainly, one per line.
It teaches one narrow thing. An embedding router compares a question with
the description of every capability it can reach. People do not write the
way capabilities are described — they hedge, they apologise, they ask two
things in one breath, they name what they do not want. This data pairs
the first… See the full description on the dataset page: https://huggingface.co/datasets/Gramscii-IT/semantic-repair-routing.modality-routing-dataset
Modality Routing Dataset
This dataset materializes the dynamic modality routing data builder used by the local
mmBERT-32K modality router training pipeline. The export is intended for review,
versioning, and uploading to a Hugging Face dataset repository.
Labels
Label
ID
Description
AR
0
Text-only requests that should route to an autoregressive LLM.
DIFFUSION
1
Image-generation requests that should route to a diffusion model.
BOTH
2
Requests that benefit… See the full description on the dataset page: https://huggingface.co/datasets/llm-semantic-router/modality-routing-dataset.hobbylm-routing-dynamics
HobbyLM-1B Routing Dynamics
Router activation data captured from 20 checkpoints spanning a 100B-token sparse
Mixture-of-Experts pretraining run, with fully documented data provenance for
every input token. Base model: harims95/hobbylm-1b-hf.
What this is
For each of 20 training checkpoints, the model was run over an identical frozen
probe set and its router's per-token, per-layer decisions were recorded: which
experts were selected, the router's raw scores before… See the full description on the dataset page: https://huggingface.co/datasets/harims95/hobbylm-routing-dynamics.meta-routing
MetaRouting Dataset
This dataset contains synthetic benchmark artifacts for the Research MetaRouting project, covering meta-decision policies for agentic workflows: when to answer directly, decompose, retrieve, execute code, delegate, verify, or recover from failures.
Source repository: https://github.com/anote-ai/Research-MetaRouting
Displayable Configs
The Hugging Face viewer reads normalized JSONL tables under viewer/:
dai2026_traces, dai2026_tasks… See the full description on the dataset page: https://huggingface.co/datasets/anote-ai/meta-routing.prompt-routing-en
Prompt Routing — English
189,422 rows. English counterpart to
oddadmix/arabic-prompt-routing
— same recipe, not a translation of the Arabic rows. Each row is a text, a set of free-text
categories, and which one it belongs to. Categories are arbitrary English — the point is a
model that routes into a label set it has never seen.
Same task as LiquidAI/LFM2.5-Encoder-350M-Prompt-Router.
split
rows
what it tests
train
173,223
unseen_lanes
2,707
new category sets… See the full description on the dataset page: https://huggingface.co/datasets/oddadmix/prompt-routing-en.dataset-A-routing-eval
Dataset A - Routing Evaluation
Total rows: 5339 | Gated (masked) rows: 0
Stratified dataset to evaluate 3 LLM models and 3 routing systems across 6 capabilities.
Multi-config layout
from datasets import load_dataset
# Full dataset
ds = load_dataset("massaindustries/dataset-A-routing-eval", "all")
# Per dimension
ds_math = load_dataset("massaindustries/dataset-A-routing-eval", "math_reasoning")
ds_code = load_dataset("massaindustries/dataset-A-routing-eval", "coding")
#… See the full description on the dataset page: https://huggingface.co/datasets/massaindustries/dataset-A-routing-eval.kairos-routing
Kairos Routing
Kairos Routing is a long-format dataset for training a model to select the
best AI model for a prompt. Each row describes one candidate model evaluated
on one prompt.
The intended learning problem is:
prompt + candidate_model_vector -> score
The router can score several candidate models for the same prompt and select
the model with the highest predicted score.
Dataset Summary
Approximately 1.92 million rows
Approximately 144,000 unique prompts
64… See the full description on the dataset page: https://huggingface.co/datasets/sijirama/kairos-routing.Prompt-Routing-Dataset-ko
Prompt-Routing-Dataset-ko
Korean translation of SupraLabs/Prompt-Routing-Dataset.
992 rows.
Only the prompt column was translated. Every label was copied verbatim. The English
source is kept as prompt_en so any row can be audited.
Columns
column
type
prompt
string
Korean translation
prompt_en
string
English source, copied unchanged
complexity_score
int
1–5
primary_domain
string
free text, 338 distinct values
coding_task
bool
math_task
bool… See the full description on the dataset page: https://huggingface.co/datasets/atonlee/Prompt-Routing-Dataset-ko.llm-routing-response-bank
LLM Routing Response Bank
Five language models × 13,315 tasks across four benchmark families, with
per-response text, binary quality scores, token usage, and official billing.
Collected for a routing study with a paired calibration/evaluation design:
256 calibration tasks, 13,059 evaluation tasks.
Contents
file
rows
note
tasks_cal.jsonl / tasks_eval.jsonl
256 / 13,059
prompts + reference answers; gpqa_diamond rows are hash-only (see below)… See the full description on the dataset page: https://huggingface.co/datasets/Lurume/llm-routing-response-bank.aml-routing-benchmark-v0
AML Routing Benchmark v0 — Free Eval Kit
Free small synthetic modular-routing benchmark (~300 records across train/eval/hard + source-backed subset).
Released on the Hub as a marketing / eval kit. Optional paid packaging may exist elsewhere; this Hub copy is free under the same no-redistrib license.
What’s inside
data/v0.4/train.jsonl (131)
data/v0.4/eval.jsonl (37)
data/v0.4/hard.jsonl (72)
data/v0.5/source_backed_60.jsonl (60)
License
Custom… See the full description on the dataset page: https://huggingface.co/datasets/Primitive-Origins/aml-routing-benchmark-v0.llm-routing-text-classification
Prompt Task Clasification
Category prompt into categories and results into the most probably task
Current Supported Categories
['fill_mask',
'conversation',
'midjourney_image_generation',
'math',
'science',
'toxic_harmful',
'logical_reasoning',
'sex',
'creative_writing']
Categories Data Composition
![Categories Composition](data:image/png;base64… See the full description on the dataset page: https://huggingface.co/datasets/jeanvydes/llm-routing-text-classification.repro-r2eval-routing-eval-bundle
R²Eval reproduction bundle
Reduced-scale, real reproduction of the routing claims of ICML 2026 paper
"Routing and Reasoned Evaluation with Large Language Models" (R²Eval)
(OpenReview d0dDhLR19Y, submission #9427).
What the paper claims
Claim 1: a routing-aware automated-assessment (LLM-as-judge) framework reduces evaluation
cost and latency while keeping alignment with human assessment.
Claim 2: difficulty-aware offline/online routing yields substantially better… See the full description on the dataset page: https://huggingface.co/datasets/kpshinnik/repro-r2eval-routing-eval-bundle.dataset-A-routing
Dataset A - Routing (3 modelli, verdict-level)
Totale query: 5504 | Gated (masked) query: 0
Dataset di valutazione per LLM routing systems su 6 capability. Ogni query è stata eseguita su 3 modelli (qwen3.5-9b, deepseek-v4-flash, kimi2.6) e giudicata con grader deterministici (math/coding/ifeval) o LLM judge panel 2-of-3 (planning_agentic) / single judge (creative_synthesis, world_knowledge).
Configs
from datasets import load_dataset
# Pivot verdict per query (default… See the full description on the dataset page: https://huggingface.co/datasets/massaindustries/dataset-A-routing.support-routing-ml-20260910-dataset
Applied ML Support Router Synthetic Dataset
Summary
This dataset contains 14 training examples and 4
held-out examples for Support operations need reproducible routing models that expose confidence and defer uncertain cases.
Every record is synthetic and includes:
input: query, event, or feature description
label: expected class, route, relation, or evidence category
context: synthetic supporting context
source: fictional source identifier
variant: generation… See the full description on the dataset page: https://huggingface.co/datasets/RKB109/support-routing-ml-20260910-dataset.russian-medical-routing-dataset
Russian Medical Routing Dataset
Назначение
Датасет предназначен для классификации русскоязычных жалоб пациентов по медицинским специализациям.
Состав
Всего записей: 9988
Self-Instruct: 2988
Confidence sampling: 3000
Hybrid: 4000
Получение данных
LLM: google/gemini-2.5-flash-lite через OpenRouter.
Self-Instruct: генерация новых жалоб по заданной специализации.
Confidence sampling: 5 независимых классификаций; confidence определяется… See the full description on the dataset page: https://huggingface.co/datasets/petaevd/russian-medical-routing-dataset.multilingual-routing-decodesPrompt-Routing-Dataset-zh-cnPrompt Routing Dataset · Multi-Task Infrastructure Routing
Only the prompt was translated into Chinese.
About this dataset
This dataset is a highly dense, premium alignment asset explicitly designed to train Edge Orchestrators and Routing Models ranging from 50M to 1.5B parameters.
When deploying small language models (SLMs) on consumer hardware or local edge instances, running multi-step mathematical derivations or complex architectural software tasks often causes catastrophic… See the full description on the dataset page: https://huggingface.co/datasets/aifeifei798/Prompt-Routing-Dataset-zh-cn.deniz-unay-v6-expert-routing
Deniz UNAY V6 — Expert Routing Dataset
Overview
Deniz UNAY V6 is an evidence-oriented expert-routing and intent-classification dataset designed to evaluate whether a user query should be routed to a specific non-clinical domain expert.
The dataset focuses on expert matching for:
Technology addiction awareness
Digital addiction education
Screen-time management
Digital detox
Digital literacy
Cyberbullying awareness
Digital wellbeing
Parent education
Teacher… See the full description on the dataset page: https://huggingface.co/datasets/MrDen1234567890/deniz-unay-v6-expert-routing.perceive-benchmark
PERCEIVE
PERCEIVE (Psychophysics-driven Elicitation for Routing Cost-Efficiency In
Vision-Language Evaluation) is a 4,801-sample document-image QA benchmark for
cost-aware VLM routing. Each sample carries psychophysical complexity annotations
(Visual Dependency Score, Reasoning Depth Score, Spatial Extent Score) and a
routing label identifying the cheapest model-budget configuration that answers
it correctly.
Routing labels are derived via a QUEST-style adaptive cascade achieving… See the full description on the dataset page: https://huggingface.co/datasets/quantiphi-routing/perceive-benchmark.llm-routing-attack-data
MauroPello/llm-routing-attack-data
This dataset contains the JSONL splits used for LLM routing attack experiments.
The Hub exposes the full dataset as the default configuration and the smaller sample as the reduced configuration when both are present.
Files
File
Rows
Size
full/train.jsonl
68687
63.1 MB
full/val.jsonl
14721
13.5 MB
full/test.jsonl
14721
13.5 MB
reduced/train.jsonl
4382
4.1 MB
reduced/val.jsonl
941
906.4 KB
reduced/test.jsonl… See the full description on the dataset page: https://huggingface.co/datasets/MauroPello/llm-routing-attack-data.uai-orb-synthetic-routing
UAI Orb synthetic routing labels
Design-hypothesis labels (SOLO / COUNCIL / RELAY) for synthetic and golden eval prompts. No user conversations. Not a human gold set.
Live chat (2026-09-05) uses an explicit Orb menu including RACE. RACE is not a label in this dataset. Production routing is Cloudflare Workers AI incumbents (GLM-5.3-flash SOLO, GLM-5.3 agent, DeepSeek V4 Pro / Kimi K2.7 satellites). This classifier stays off (FEATURE_ORB_ROUTER).