Datasets:

James-Cuda
/

tinyrouter-m1

pretty_name: tinyrouter-m1
task_categories:
  - text-classification
tags:
  - tinyrouter
  - milestone-1
  - gci-bench
license: other
configs:
  - config_name: 5-domain
    data_files:
      - split: train
        path: 5-domain/train.parquet
      - split: test
        path: 5-domain/test.parquet
  - config_name: 20-domain
    data_files:
      - split: train
        path: 20-domain/train.parquet
      - split: test
        path: 20-domain/test.parquet
  - config_name: default
    data_files:
      - split: train
        path: 5-domain/train.parquet
      - split: test
        path: 5-domain/test.parquet

tinyrouter-m1

Two separate Milestone-1 triage datasets. Same schema; different domain taxonomies.

Overview

config	#domains	train	test	total	difficulty levels present
`5-domain`	5	311,868	16,414	328,282	1, 2, 3, 4, 5
`20-domain`	20	4,750	250	5,000	2, 3, 4

Schema (both)

column	type	meaning
`id`	string	stable row id (`source:split:idx`)
`domain`	string	closed domain label for that config
`difficulty`	int	1–5
`prompt`	string	user / problem text

`5-domain` — detailed

Full Milestone-1 pool remeshed to 5 router buckets.

Difficulty: 1=118, 2=92,844, 3=218,087, 4=14,812, 5=2,421

domain	description / sources	train	test	total
`math`	GSM8K, MATH-500, AIME, AQuA-RAT, …	102,658	5,442	108,100
`code`	HumanEval, MBPP, BigCodeBench, …	3,009	158	3,167
`knowledge`	MMLU, MMLU-Pro, GPQA, ARC, TruthfulQA, GCI, …	134,537	7,109	141,646
`commonsense`	HellaSwag	57,027	2,923	59,950
`instruction`	Dolly-15k, IFEval	14,637	782	15,419

`20-domain` — detailed

GCI-Bench rows only (GPL-3.0). domain = official GCI topic; topic labels from upstream topicLabel. Difficulty: easy→2 / medium→3 / hard→4.

Difficulty: 2=1,447, 3=2,278, 4=1,275

domain	topic label	train	test	total	difficulty (2/3/4)
`agriculture`	Agriculture & Crops	228	22	250	75/120/55
`archaeology`	Archaeology	237	13	250	69/121/60
`architecture`	Architecture & Construction	237	13	250	76/121/53
`astronomy`	Astronomy & Space	242	8	250	69/115/66
`automotive`	Automotive Mechanics	241	9	250	60/130/60
`aviation`	Aviation	241	9	250	74/117/59
`chemistry`	Chemistry Lab Processes	242	8	250	72/114/64
`cooking`	Culinary Science	241	9	250	82/105/63
`energy`	Renewable Energy	232	18	250	81/103/66
`finance`	Personal Finance	239	11	250	74/104/72
`gardening`	Home Gardening	235	15	250	75/110/65
`hardware`	Computer Hardware	237	13	250	66/113/71
`marine`	Marine Biology	237	13	250	87/94/69
`music`	Music Theory & Instruments	239	11	250	82/113/55
`photography`	Photography	236	14	250	60/111/79
`physiology`	Human Physiology	235	15	250	66/113/71
`sports`	Sports Training	233	17	250	76/120/54
`textiles`	Textile & Fashion	245	5	250	75/108/67
`weather`	Weather & Climate	239	11	250	70/109/71
`wildlife`	Wildlife & Animal Behavior	234	16	250	58/137/55

Load

from datasets import load_dataset
ds5  = load_dataset("James-Cuda/tinyrouter-m1", "5-domain", split="train")
ds20 = load_dataset("James-Cuda/tinyrouter-m1", "20-domain", split="train")
print(ds5[0]["domain"], ds5[0]["difficulty"])
print(ds20[0]["domain"], ds20[0]["difficulty"])

Built by scripts/build_tinyrouter_m1_slim.py.