RepoJournal
Alignment & Post-training Alignment & Post-training
24 wires and counting

$ follow Alignment & Post-training

Keep up with Alignment & Post-training in about 3 minutes: what actually shipped — the commits, pull requests, releases, and security advisories that matter.

or

fair warning: these emails are deeply technical. diffs, version numbers, CVEs, benchmark deltas. if that's not your idea of a good read, this isn't your newsletter.

Folds into your digest — weekly by default, monthly if you prefer. Unsubscribe in one click.

$ status

wire 2026-10-29
stories 6

© 2026 RepoJournal Home Showcase How it works Privacy

$ the-wire · showcase

MILES evaluations move to native SGLang; compiled attention base drops install layer to 2.77 MB

By RepoJournal · Filed · About Alignment & Post-training · Composed from the cited sources · methodology

open-instruct landed two MILES runtime changes that alter how background evals launch and how the attention-kernel image is built, plus a README refresh that swaps the TÜLU 3 launchers for the maintained ones.

Use native SGLang for MILES background evaluations allenai/open-instruct

by no0p

MILES background policy evaluations now select olmo-eval's native `sglang` provider and the Olmo registration hook, and `node.toml` moves from shared-engine evaluation to a separate one-GPU GSM8K/GSM8K-Platinum/IFBench panel running every 50 optimizer updates plus at the start and end, while `three-node.toml` keeps its full 25-entry panel and dedicated safety-judge GPUs. Operators should note t...

Pin the compiled MILES base and layer runtime sources allenai/open-instruct

by no0p

The published compiled MILES base now bundles FA2, FA3 and FA4 as locked sources, so application builds install attention kernels instead of recompiling them. The per the PR, the earlier 8.06 GB install layer shrank to 2.77 MB with the cache mount, and builds pin to `robertb/miles-binary-base-20261027` (Docker digest `sha256:947e615f93cb88f8739ba270f30adc7f012af8bc0abf43e96f11c`).

Remove stale and redundant README content (#1971) allenai/open-instruct

by Abhishek Rao

The README now documents the maintained SFT and DPO launchers in place of the misleading TÜLU 3 examples, which means the commands you copy from the front page are the ones that still work. The old news feed, duplicated checkpoint table, outdated image instructions and redundant pre-commit dependency step are gone.

Remove stale and redundant README content allenai/open-instruct

by abhishekraok

The companion PR to the README cleanup restores a current DPO entry point, adds SFT launch commands and rebuilds the repo tree section, while leaving the title, logo, licensing text, acknowledgements and citations untouched.

Quick answers

What shipped in Alignment & Post-training on October 29, 2026?
open-instruct landed two MILES runtime changes that alter how background evals launch and how the attention-kernel image is built, plus a README refresh that swaps the TÜLU 3 launchers for the maintained ones. In total, 3 commits and 3 pull requests landed.
Who contributed to Alignment & Post-training on October 29, 2026?
2 developers shipped this update, including no0p and Abhishek Rao.
What were the notable Alignment & Post-training updates?
Use native SGLang for MILES background evaluations, Pin the compiled MILES base and layer runtime sources, and Remove stale and redundant README content (#1971).