Who contributed to Alignment & Post-training on October 25, 2026?

1 developer shipped this update, including Robert Berry.

♥

Oct 25, 2026

$ the-wire · showcase

GSM8K scoring parses answers as exact decimals

By RepoJournal · Filed 06:59 UTC on October 25, 2026 · About Alignment & Post-training · Composed from the cited sources · methodology

1 person shipped this

Robert Berry @no0p 1 cited

allenai/open-instruct no longer zeroes out GSM8K responses whose final number is written with trailing zeros, so any eval numbers you have on record for that benchmark are worth re-reading.

Compare GSM8K answers as exact numbers (#1921) allenai/open-instruct

by Robert Berry

GSM8KVerifier compared the last number in a response against the label as strings, so a correct 96.00 scored zero against a label of 96. It now parses both sides with Decimal after stripping thousands separators, which means 96.00 == 96 and 42 == 42.0, while 96.50 != 96 and 0.33 != 0.3333 with no tolerance; non-numeric labels keep the existing case-insensitive string match. The pull request not...

Action items

→ Re-run evals that include a GSM8K score before this fix and compare against your stored results allenai/open-instruct [plan]

$ ls alignment/reviews/ # the step-back reads

the week in review verl breaks out unified QAT and low-precision rollout

Full archive → Weekly reviews →

Quick answers

What shipped in Alignment & Post-training on October 25, 2026?: allenai/open-instruct no longer zeroes out GSM8K responses whose final number is written with trailing zeros, so any eval numbers you have on record for that benchmark are worth re-reading. In total, 1 commits and 1 pull requests landed.
Who contributed to Alignment & Post-training on October 25, 2026?: 1 developer shipped this update, including Robert Berry.
What were the notable Alignment & Post-training updates?: Compare GSM8K answers as exact numbers (#1921).

♥

$ log Alignment & Post-training --recent

2026-10-24 MILES ships a 25-benchmark eval panel and simpler config presets 10 shipped 2026-10-23 Open-Instruct moves to Miles v0.1.2; verl fixes teacher top-k packing under CP 24 shipped 2026-10-22 verl standardizes low-precision rollout kernels on DeepGEMM 20 shipped 2026-10-21 TPU GRPO image fixed, Megatron optimizer offload leaves HDO state alone 24 shipped

$ log --all # elsewhere on the wire

AI Red Teaming 2026-10-25 PyRIT closes a judge-spoofing hole, promptfoo accepts Langfuse v6 Safety Evals 2026-10-25 Viewer plays terminal sessions, Kimi K3.1 Pro lands in 0.3.284

$ ls showcases/

$ init your-desk

RepoJournal can write one of these for any repo you care about. Sign in with GitHub, pick your repos and sources, and your desk gets a wire like this one every morning.