verify_slime_vendor.sh proves frameworks/slime is still slime @ 26525f9.
verify_launcher_flags.sh proves every flag the launcher passes is declared by slime, Megatron, or SGLang's prefixed ServerArgs -- a typo'd flag is not a crash, the setting is simply dropped.
SciWorld OPD on slime: launcher rewritten to match upstream, 1+1 GPU path, data prep
The launcher now mirrors reopd_ref scripts/train/opd_math.sh structurally (Ray, --colocate, SGLANG_ARGS, MISC_ARGS, full OPTIMIZER_ARGS) instead of calling train.py directly. The old path never passed MODEL_ARGS, and slime takes the student's architecture from those flags, so it could not have produced a comparable run.
Added: env_1gpu.sh / smoke_1gpu.sh / run_1gpu.sh for one student GPU and one teacher GPU; prepare_sciworld_tasks.py, which writes PROMPT_DATA and a games_fingerprint.json (item_id is a position in a list rebuilt from the installed scienceworld package -- a different version renumbers it and nothing raises); probe_teacher_temperature.py, which settles empirically whether SCIWORLD_TEACHER_TEMPERATURE reaches the returned logprobs at all, since sampling params normally shape generation and not the scoring of tokens already present.
Train is every game index not in the eval set, so the two cannot overlap. AgentGym's own SciWorld split leaks 61 of its 200 test ids into tra…
Vendor slime @ 26525f9, the trainer reopd_ref builds on (part 2)
Byte-identical copy of reopd_ref/third_party/slime (BaohaoLiao/slime, commit 26525f97179afc47c9ef6fff045dbc28d79f1875). Pinning it means a run here does not depend on whatever `pip install slime` resolves to today: a silently drifted trainer does not raise, it just stops being the baseline everyone else reports.
Includes reopd_scripts/models/*.sh (upstream MODEL_ARGS, also byte-identical) plus qwen3-1.7B.sh, which upstream does not ship and which a one-GPU student needs, and opd_math.sh.upstream to diff the launcher against.
Vendor slime @ 26525f9, the trainer reopd_ref builds on
Byte-identical copy of reopd_ref/third_party/slime (BaohaoLiao/slime, commit 26525f97179afc47c9ef6fff045dbc28d79f1875). Pinning it means a run here does not depend on whatever `pip install slime` resolves to today: a silently drifted trainer does not raise, it just stops being the baseline everyone else reports.
Includes reopd_scripts/models/*.sh (upstream MODEL_ARGS, also byte-identical) plus qwen3-1.7B.sh, which upstream does not ship and which a one-GPU student needs, and opd_math.sh.upstream to diff the launcher against.
