RepoJournal
Alignment & Post-training Alignment & Post-training
29 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 2026-11-01
stories 10

© 2026 RepoJournal Home Showcase How it works Privacy

$ the-wire · showcase

open-instruct reworks MILES defaults, fixes reward-hacking IFEvalG checkers

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

allenai/open-instruct changed MILES runtime defaults and the batch handling around them, and separately fixed three IFEvalG checkers that were paying reward for degenerate answers.

Consolidate MILES throughput configuration, runtime and evaluation optimizations allenai/open-instruct

by no0p

MILES now uses the Rust router by default, runs distributed HTTP workers, and drops the successful-refresh admission handshakes while keeping engine pause/update/resume; evaluation picks model/GPU profiles with capacity-aware batches. Every one of those is overridable, so existing explicit configs keep working, but anyone relying on the old defaults will see different throughput and eval batchi...

Fix IFEvalG checkers that rewarded degenerate answers allenai/open-instruct

by abhishekraok

Three IFEvalG checkers paid reward for degenerate answers or paid at random, and both the GRPO IFEvalVerifier and the MILES ManifestIFVerifier delegate to them. If you trained RL on the IF pool with these checkers, the reward signal behind those runs was wrong, and the table in the pull request maps each constraint name to its bug and fix.

Improve MILES batch delivery and align demo learning defaults allenai/open-instruct

by no0p

Batches were delayed by rollout delivery and repeated trainer preparation; MILES now pins a revision that drains ready groups, writes sampled diagnostics asynchronously and reuses prepared token packs. It also exposes an opt-in core.layer_profile_interval if you want the per-layer, expert and cache cost measurements, and the demos now share per-response loss, zero truncation rewards and eight r...

[ci] chore: enable daily Ascend A5 Docker builds (#8470) verl-project/verl

by lxb007981

verl turned on the existing cron so Ascend A5 Docker images build daily at 19:00 UTC, matching the Ascend A3 workflow, which matters only if you pull those images to run on Ascend hardware. The README hardware support entry now names the vendor as "Huawei Ascend" rather than just "Ascend."

Quick answers

What shipped in Alignment & Post-training on November 1, 2026?
allenai/open-instruct changed MILES runtime defaults and the batch handling around them, and separately fixed three IFEvalG checkers that were paying reward for degenerate answers. In total, 5 commits and 5 pull requests landed.
Who contributed to Alignment & Post-training on November 1, 2026?
4 developers shipped this update, including lxb007981, XiangRun Liu, no0p, and abhishekraok.
What were the notable Alignment & Post-training updates?
Consolidate MILES throughput configuration, runtime and evaluation optimizations, Fix IFEvalG checkers that rewarded degenerate answers, and Improve MILES batch delivery and align demo learning defaults.