mirror of
https://github.com/wassname/evil_MoE.git
synced 2026-07-10 01:54:42 +08:00
chore: justfile build-substrate + run-substrate recipes
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
@@ -129,6 +129,23 @@ run-cell-mode ENVMODE SEED:
|
||||
--steps=60 --seed={{ SEED }} \
|
||||
--out-tag=_emerge_{{ ENVMODE }}_s{{ SEED }}
|
||||
|
||||
# Build the even, non-overlapping multi-loophole teacher batch (substrate) from the
|
||||
# de-risk elicit-then-strip hacks + the run_tests teacher pool. Writes
|
||||
# out/pools/substrate/{prompt_*.jsonl.gz, partition.json}. Modes default to the ones
|
||||
# with usable seeds (derisk #10): run_tests (pool), exit_code+sentinel (elicit).
|
||||
build-substrate MODES="run_tests,exit_code,sentinel":
|
||||
uv run python -m projected_grpo.build_substrate \
|
||||
--modes {{ MODES }} --pool-modes run_tests --min-hacks 5
|
||||
|
||||
# Vanilla-GRPO emergence on the multi-loophole substrate: does the student learn ALL
|
||||
# K loopholes from the repeated even teacher batch? UAT = end-of-run SUBSTRATE table
|
||||
# (per-mode hacks>0 + finite first_step). v_hack is loaded only for the cos diagnostic.
|
||||
run-substrate INTERV="none" SEED="41" MIX="0.25" STEPS="80":
|
||||
{{ TRAIN }} fast --intervention={{ INTERV }} \
|
||||
--teacher-pool-dir=out/pools/substrate --mix-ratio={{ MIX }} \
|
||||
--v-hack-path=out/vhack/v_hack_21pairs.safetensors \
|
||||
--seed={{ SEED }} --steps={{ STEPS }} --out-tag=_substrate3_{{ INTERV }}_s{{ SEED }}
|
||||
|
||||
# Regenerate both dynamics plots from the cell logs (default: all cells; pass a
|
||||
# narrower glob like 'logs/*_cell_*_s41.log' for the seed-41-only checkpoint).
|
||||
regen-dynamics GLOB='logs/*_cell_*.log':
|
||||
|
||||
Reference in New Issue
Block a user