Base: google/t5-efficient-tiny (~16M params)
Shipped: INT8 ONNX (encoder + decoder + decoder-with-past), ~36 MB total
Accuracy: 99.2% on 4 000 fresh synthetic samples · 97.3% on a hand-crafted adversarial set
Validity: 100% of outputs are parseable 5-field cron
Latency: ~3 ms / item on a laptop CPU

Usage

Browser / Node — transformers.js

import { pipeline } from "@huggingface/transformers";
const cron = await pipeline("text2text-generation", "pavstev/cronlm");
const out = await cron("every weekday at 9am".toLowerCase());
console.log(out[0].generated_text); // "0 9 * * 1-5"

Lowercase the input before passing — bumps accuracy on Title-Case inputs by ~2.5 pp.

Python — transformers

from transformers import pipeline
cron = pipeline("text2text-generation", model="pavstev/cronlm")
print(cron("every weekday at 9am")[0]["generated_text"])  # "0 9 * * 1-5"

Examples

Input	Output
every 5 minutes	`/5 * * *`
every weekday at 9am	`0 9 * * 1-5`
at 10:30 PM on Sundays	`30 22 * * 0`
every 15 minutes between 9 and 17	`/15 9-17 * *`
every Mon and Wed at noon	`0 12 * * 1,3`
on the 1st of every month	`0 0 1 * *`
every quarter	`0 0 1 1,4,7,10 *`
at midnight	`0 0 * * *`
`@daily`	`0 0 * * *`
MWF at 9am	`0 9 * * 1,3,5`
every December 25	`0 0 25 12 *`
from the 1st to the 5th at noon	`0 12 1-5 * *`
every 30 minutes during business hours	`/30 9-17 * 1-5`

Validation

A pure-Python validator/fixer is included in the cronlm package — wrap model output to catch any rare invalid generation:

from cronlm import fix
fix("0 25 * * 1")  # "0 23 * * 1"

Training

Source, dataset generator, and training scripts: https://github.com/pavstev/cronlm

License

MIT