cd /news/machine-learning/2m-reasoning-examples-for-fine-tunin… · home › topics › machine-learning › article
[ARTICLE · art-173809] src=discuss.huggingface.co ↗ pub= topic=machine-learning verified=true sentiment=↑ positive

2M reasoning examples for fine-tuning small models, with worked solutions and a Python answer checker

A dataset creator released CAEL Verified Reasoning on Hugging Face, a CC BY 4.0 collection of 2 million synthetic reasoning examples for fine-tuning small language models, spanning arithmetic, logic, multi-turn memory, conflicting reports and grid puzzles. The dataset splits into 1.98 million training examples, 20,000 test examples and a separate 1,427-example hard split, with roughly 37,000 examples covering missing information, contradictory premises or facts the user never supplied. The release includes the rule-based generator and a Python answer checker so users can create fresh examples and score model replies, leaving open whether fine-tuning still helps when someone else writes the questions.

key takeaway

CAEL Verified Reasoning offers 1.98 million training examples plus a 1,427-example hard split generated by the same rule-based solvers, letting builders test whether fine-tuning generalizes to questions written by someone else.

read1 min views5 publishedNov 21, 2026

I built CAEL Verified Reasoning for people training small language models. It has 2 million synthetic examples across eight areas, including arithmetic, logic, multi-turn memory, conflicting reports, and grid puzzles. The rule-based solvers behind CAEL generate each problem, calculate its reference answer, and write out the solution. The dataset is licensed under CC BY 4.0.

One question I wanted to explore was whether a model can tell when it has enough information to answer. Say I tell you my second bike ride took four minutes less than my first and ask how long the first ride took. You’d have to ask me for another detail. If I add that the second ride took 18 minutes, you can work out the answer. Around 37,000 examples deal with missing information, contradictory premises, or facts the user never supplied.

There are 1.98 million training examples, 20,000 test examples, and a separate hard split of 1,427. I’ve included the generator and a Python answer checker so you can make fresh examples and score model replies. The hard split uses the same generator, which leaves an open question I’d like to test: does fine-tuning still help when someone else writes the questions?

You can inspect a training example without down the whole dataset:

from datasets import load_dataset

from datasets import load_dataset

rows = load_dataset("vynly/cael-verified-reasoning", split="train", streaming=True)
row = next(iter(rows))
print(row["messages"], row["answer"], row["solution"], sep="\n\n")

CAEL Verified Reasoning on Hugging Face

If you try it with a small model, I’d love to know which one you used and see one question where a change in wording made it go from right to wrong. Those examples would help me decide what to add next.

── more in #machine-learning 4 stories · sorted by recency
── more on @cael verified reasoning 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/2m-reasoning-example…] indexed:0 read:1min 2026-11-21 · —