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 2026-10-31
stories 8
RSS Archive Reviews

© 2026 RepoJournal Home Showcase How it works Privacy

$ the-wire · showcase

MILES gains fixed source mixtures, and 500s stop scoring as zero

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

abhishekraok no0p
2 people shipped this
abhishekraok @abhishekraok 3 cited no0p @no0p 1 cited

open-instruct's MILES reward stack took three operator-visible changes: fixed per-source mixtures for training runs, retries for transient HTTP 500s from the code service, and a TPU image that makes long SFT resumes dramatically faster.

Add MILES source mixtures and MIT SynLogic Sudoku allenai/open-instruct

by no0p

Runs can now reference Dolci Think intact alongside a new source via explicit positive fractions that must sum to 1, with preparation writing separate per-source files plus a hashed mixture manifest and checking prompt lengths and train/held-out overlap across sources, so you no longer build an interleaved training file by hand. Existing single-dataset configs keep the shuffle provider, but glo...

Retry transient HTTP 500s from the MILES code service before scoring zero allenai/open-instruct

by abhishekraok

The MILES code reward previously scored an HTTP 500 from the code-execution service as 0 and never retried it, because the urllib3 retry list covered only 502/503/504. In the H120 RL runs on Olmo 4T EMO base, 500s accounted for 4.0 to 4.9% of code calls per attempt and returned in 0.5 to 4.1 s; all 8 of 8 replayed 500 programs came back HTTP 200, so unfixed 500s were silently zeroing achievable...

Add a TPU image with a pooled chat-template map for fast resume allenai/open-instruct

by abhishekraok

On resume, MaxText SFT replays every already-trained batch through a serial chat-template map at roughly 1 s per skipped step for the Olmo 3.5 hero recipe, so a 56k-step run resumed late spends hours replaying before it trains. The new image-olmo35-fs image adds an opt-in spawned pool via `MAXTEXT_SFT_TEMPLATE_WORKERS`, reported to yield the same batches about 10x faster; leaving it unset keeps...

Add a TPU image with a pooled chat-template map for fast resume (#2115) allenai/open-instruct

by Abhishek Rao

The merged half of the same work lands the pooled chat-template map in `scripts/tpu/maxtext/image-olmo35-fs/`, built on the published hero image so it drops in without changing anything else in the recipe.

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

the week in review verl pins DeepEP for uv Megatron, Qwen3.5 TPU recipe the month in review verl lands score centering for bypass-mode REINFORCE
Full archive → Weekly reviews →

Quick answers

What shipped in Alignment & Post-training on October 31, 2026?
open-instruct's MILES reward stack took three operator-visible changes: fixed per-source mixtures for training runs, retries for transient HTTP 500s from the code service, and a TPU image that makes long SFT resumes dramatically faster. In total, 4 commits and 4 pull requests landed.
Who contributed to Alignment & Post-training on October 31, 2026?
2 developers shipped this update, including no0p and abhishekraok.
What were the notable Alignment & Post-training updates?
Add MILES source mixtures and MIT SynLogic Sudoku, Retry transient HTTP 500s from the MILES code service before scoring zero, and Add a TPU image with a pooled chat-template map for fast resume.
♥

$ log Alignment & Post-training --recent

MILES wall-clock stops resume by default, retained generation state in learning examples 18 shipped verl lands native NVFP4 training and rollout, open-instruct overlaps MILES startup 26 shipped FP8 rollout KV cache fix in verl, refresh batches overlap training in open-instruct 10 shipped open-instruct reworks MILES defaults, fixes reward-hacking IFEvalG checkers 10 shipped

$ log --all # elsewhere on the wire

Agent Security Judge prompts stop leaking to hooks-server logs AI Red Teaming PyRIT's GUI gets workflow defaults, quiet auto-refresh and in-repo scorers move to the expectation-aware hook

$ ls showcases/

Safety Evals AI Red Teaming Agent Security Interpretability

$ 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.