skycmd-math / README.md
maxgfr's picture
Card: the new name, and the closed experiment's verdict
dea4767 verified
|
Raw History Blame Contribute Delete
2.35 kB
metadata
license: mit
language:
  - en
pretty_name: skycmd math questions (GSM8K and MATH)
task_categories:
  - question-answering
tags:
  - skycmd
configs:
  - config_name: calib
    data_files: data/calib.parquet
  - config_name: test
    data_files: data/test.parquet

skycmd math questions (GSM8K and MATH)

The calibration and test questions of the LFM3-1.7B System 1 / System 2 router (docs/lfm3-s1s2-prereg.md): GSM8K and MATH, drawn with a fixed seed at pinned revisions. That experiment is closed and failed its pre-registered gate: the router used 1,032 tokens per question on average, where the gate required at most half of the 958 used by always thinking (H3, compute halved). This data set was named skycmd-lfm3-math until 19/11/2026; the old address redirects here.

Part of skycmd, tiny decision models trained from scratch in Rust that run in the browser. Written by scripts/lfm3-s1s2.sh data at commit 84c6d854; the script was removed from main on 19/11/2026 and is in the git history at that commit. Licences of every set: DATA_LICENSES.md.

Each row is one question: id, task (gsm8k or math), question and answer (the final answer, compared after normalization).

Sources and licences

Each config is one source and keeps its source's licence (share-alike sources stay share-alike). Attribution goes to the upstream authors; the conversion is ours.

config upstream licence how it is used rows
calib openai/gsm8k (main, train, 740312ad) and DigitalLearningGmbH/MATH-lighteval (train, 0530c786) mit 300 + 300 calibration questions, seed 1234 600
test openai/gsm8k (main, test) and HuggingFaceH4/MATH-500 (6e4ed1a2) mit 500 + 500 test questions, seed 1234 1,000

Files

file rows
data/calib.parquet 600
data/test.parquet 1,000
from datasets import load_dataset
ds = load_dataset("maxgfr/skycmd-math", "calib")

skycmd wrote no row with a language model: the targets come from the upstream labels (some upstream sets carry their own teacher soft labels, see their cards), the games' oracles or our generators.