$ the-wire · showcase
Anchored math grading by default in MILES, and one launcher for every dense Qwen3 TPU recipe
By RepoJournal · Filed · About Alignment & Post-training · Composed from the cited sources · methodology
Open Instruct changed what MILES math rewards by default and moved dataset readiness out of the trainer, while verl folded six dense Qwen3 TPU scripts into one launcher and added a TPU MoE recipe.
Use anchored Math-Verify grading by default in MILES allenai/open-instruct
MILES general math grading now runs pinned Hugging Face Math-Verify 0.9.0 with anchored prediction extraction: answers must be boxed or follow a recognized answer phrase, because unrestricted extraction was found to reward incidental numbers in unfinished responses. If you run the math route, expect scores to shift and check any replies that relied on unboxed final answers.
[hardware, recipe] feat: unify dense Qwen3 (0.6B to 32B) TPU GRPO recipes into run_qwen3_torchtitan.sh verl-project/verl
The dense Qwen3 TPU GRPO scripts in examples/tpu/grpo/ collapse into a single launcher, run_qwen3_torchtitan.sh, selected with MODEL_SIZE=0.6B | 1.7B | 4B | 8B | 14B | 32B. It replaces run_qwen3_0_6b_torchtitan.sh and run_qwen3_4b_torchtitan.sh, so any job or script invoking those two files by name needs repointing; 0.6B remains the default 5-step smoke test on two v6e-1 slices.
Retire dataset readiness tooling now owned by RL Data Workbench allenai/open-instruct
The dataset screening and qualification tooling under scripts/data/rl_readiness/ is gone from Open Instruct because the RL Data Workbench owns it now; source adapters, catalog resolution, verifiers and the loader stay for run time. Any pipeline that called qualify.py or its siblings directly must move that step to Workbench.
[hardware, recipe] feat: Qwen3-30B-A3B (MoE) GRPO recipe for TPU verl-project/verl
A TPU GRPO recipe for Qwen3-30B-A3B (128 experts, 8 active) ships as run_qwen3_30b_a3b_torchtitan.sh, running TorchTitan training and vLLM rollout over two TPU slices with tpu-sync weight sync, defaulting to TPU v7x 2x2x1 plus 2x2x1, with v5p 2x2x4 covered in the README.
Validate opt-in policy-scoped prefix cache configuration allenai/open-instruct
Setting miles.sglang_radix_cache_backend="olmo_policy_versioned_v1" now gets its configuration validated, letting long-context RL share prefill across the several responses generated per prompt while recipes with the cache off behave as before; SGLang itself is still pinned and unmodified.
Action items
