mirror of
https://github.com/wassname/evil_MoE.git
synced 2026-08-19 06:20:25 +08:00
reorg: out/ sorted by datatype (vhack/ pools/ runs/ vhack_grads/ figs/)
Code writes+reads the new scheme; migrate_out_dirs.py moved 225 loose artifacts (0 left at top level). Per-run checkpoints+rollouts now group under runs/<ts>_<run_id>/ as train.safetensors/rollouts.jsonl. Figures land in out/figs/ with a stable docs/figs/<name>.png symlink (figs.link_latest). justfile also gains run-cell REFRESH param (online-erasure arm). Smoke + smoke-vanilla + results all green on new paths. Requeue manifest preserves the why/resolve labels that pueue reset wiped. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
wassname
co-authored by
Claude Opus 4.8
Claude Opus 4.8
parent
4fb7b59548
commit
4621488cc0
@@ -0,0 +1,81 @@
|
||||
# Requeue manifest (queue reset 2026-05-30)
|
||||
|
||||
Durable copy of the pueue why/resolve labels + commands, captured before
|
||||
`pueue reset` wiped them. This file is the requeue plan. Raw JSON snapshot:
|
||||
`/tmp/claude-1000/queue_snapshot_033450.json` (not durable — this .md is).
|
||||
|
||||
## Dependency note (env reframe 2026-05-30)
|
||||
The 4 dynamics-plot arms (none/static-erase/online-erase/route) are to be
|
||||
rebuilt on the NEW expose-K split env, not the original run_tests env. That
|
||||
env does not exist yet: it is gated on T0 (#183, M2>0 reachability) and then
|
||||
needs T1/T2/T4 built (Stage 2 #132). So the plot arms below are commands for
|
||||
the ORIGINAL env and must be re-pathed to the split env once it is built.
|
||||
Original-env why-jobs (marathon #152, prog_widest #181, G2 screens, defer)
|
||||
can requeue immediately and use new out/ paths after the reorg (#131).
|
||||
|
||||
### #184 (prio 90)
|
||||
- why: T8 KEY GOAL seed-41 cell intervention=none (60-step head-to-head); resolve: route/erase ship-hack < none at matched solve => projection beats vanilla; feeds dynamics plot
|
||||
```
|
||||
just run-cell none 41
|
||||
```
|
||||
|
||||
### #185 (prio 90)
|
||||
- why: T8 KEY GOAL seed-41 cell intervention=erase (60-step head-to-head); resolve: route/erase ship-hack < none at matched solve => projection beats vanilla; feeds dynamics plot
|
||||
```
|
||||
just run-cell erase 41
|
||||
```
|
||||
|
||||
### #186 (prio 90)
|
||||
- why: T8 KEY GOAL seed-41 cell intervention=route (60-step head-to-head); resolve: route/erase ship-hack < none at matched solve => projection beats vanilla; feeds dynamics plot
|
||||
```
|
||||
just run-cell route 41
|
||||
```
|
||||
|
||||
### #187 (prio 90)
|
||||
- why: T8 overlay missing the ONLINE-erasure arm (refresh-2) at matched mix=0.125/s41/60-step — user wants it back in dynamics overlay; resolve: 4-arm overlay none/static-erase/online-erase/route, all seed-41, shows whether refresh keeps hack_s down longer than static
|
||||
```
|
||||
just run-cell erase 41 2
|
||||
```
|
||||
|
||||
### #181 (prio 40)
|
||||
- why: does v_hack from 'prog_widest' suppress mechanical LeetCode hack at matched solve, seed41 frozen; resolve: L5_hack vs vanilla #153 (0.664), prog_wide #156 (0.500)
|
||||
```
|
||||
uv run python -m projected_grpo.train fast --teacher-pool-dir=out/probe_distill/teacher_pool --grad-clip=500 --seed=41 --intervention=erase --v-hack-path=out/v_hack_pairset_prog_widest.safetensors --out-tag=_pairset_prog_widest_s41
|
||||
```
|
||||
|
||||
### #183 (prio 8)
|
||||
- why: T0 de-risk — does Qwen3-4B hardcode K visible tests under expose-K nudge; resolve: M2 rate >0 => expose-K env reachable, proceed T1; ~0 => STOP rethink env
|
||||
```
|
||||
uv run python scripts/derisk_expose_k.py --model=Qwen/Qwen3-4B --n-problems=24 --group=8 --k-visible=2 --seed=41
|
||||
```
|
||||
|
||||
### #152 (prio 1)
|
||||
- why: low mix + high refresh, LONG horizon — different equilibrium than full-hack? (v_hack_21pairs, refresh-2 k=12, mix0125, s42, 1000 steps); resolve: hack_s<~0.7 and/or solve_s>~0.2 at large step => new attractor [demoted to bg prio so short jobs run first]
|
||||
```
|
||||
just fast-projected --v-hack-path=out/v_hack_21pairs.safetensors --vhack-refresh-every=2 --seed=42 --steps=1000 --out-tag=_equilib_refresh2_k12_mix0125_1000_s42
|
||||
```
|
||||
|
||||
### #137 (prio 0)
|
||||
- why: G2 screen rl-rewardhacking-leetcode-gt-monitor-screening-s65 (5-prompt batch); resolve: per-checkpoint E/C/D% table to find non-rh-s65 detector clusters
|
||||
```
|
||||
just pregen-teacher-alt ariahw/rl-rewardhacking-leetcode-gt-monitor-screening-s65 teacher_pool_rl-rewardhacking-leetcode-gt-monitor-screening-s65 5
|
||||
```
|
||||
|
||||
### #138 (prio 0)
|
||||
- why: G2 screen rl-rewardhacking-leetcode-judge-monitor-screening-s65 (5-prompt batch); resolve: per-checkpoint E/C/D% table to find non-rh-s65 detector clusters
|
||||
```
|
||||
just pregen-teacher-alt ariahw/rl-rewardhacking-leetcode-judge-monitor-screening-s65 teacher_pool_rl-rewardhacking-leetcode-judge-monitor-screening-s65 5
|
||||
```
|
||||
|
||||
### #139 (prio 0)
|
||||
- why: G2 screen rl-rewardhacking-leetcode-probe-monitor-screening-s65 (5-prompt batch); resolve: per-checkpoint E/C/D% table to find non-rh-s65 detector clusters
|
||||
```
|
||||
just pregen-teacher-alt ariahw/rl-rewardhacking-leetcode-probe-monitor-screening-s65 teacher_pool_rl-rewardhacking-leetcode-probe-monitor-screening-s65 5
|
||||
```
|
||||
|
||||
### #173 (prio -10)
|
||||
- why: does Qwen3.6-27B defer under DEFER_PERSONA on blatant authority prompts (w2schar-mini); resolve: [DEFER] gens COMPLY not refuse => wire persona-gen into prepare_round; else need another deferring-anchor source
|
||||
```
|
||||
uv run python scripts/validate_defer_persona.py
|
||||
```
|
||||
|
||||
@@ -27,20 +27,20 @@ results:
|
||||
# zero-variance bails every step, leaving the loss path uncovered.
|
||||
smoke *ARGS:
|
||||
BEARTYPE=1 CUDA_VISIBLE_DEVICES= {{ TRAIN }} smoke --intervention=erase \
|
||||
--v-hack-path=out/v_hack_smoke.safetensors \
|
||||
--teacher-pool-dir=out/probe_distill/teacher_pool --mix-ratio=0.5 {{ ARGS }}
|
||||
--v-hack-path=out/vhack/v_hack_smoke.safetensors \
|
||||
--teacher-pool-dir=out/pools/teacher_pool --mix-ratio=0.5 {{ ARGS }}
|
||||
|
||||
smoke-vanilla *ARGS:
|
||||
BEARTYPE=1 CUDA_VISIBLE_DEVICES= {{ TRAIN }} smoke --intervention=none \
|
||||
--teacher-pool-dir=out/probe_distill/teacher_pool --mix-ratio=0.5 {{ ARGS }}
|
||||
--teacher-pool-dir=out/pools/teacher_pool --mix-ratio=0.5 {{ ARGS }}
|
||||
|
||||
# Routing path: parks the hack-ward grad in delta_S_hack, ablates at eval.
|
||||
# Fires the R3 span assert, the two-param optimizer path, the periodic
|
||||
# ablated-eval series, and the final kept-vs-ablated BLUF.
|
||||
smoke-route *ARGS:
|
||||
BEARTYPE=1 CUDA_VISIBLE_DEVICES= {{ TRAIN }} smoke --intervention=route \
|
||||
--v-hack-path=out/v_hack_smoke.safetensors \
|
||||
--teacher-pool-dir=out/probe_distill/teacher_pool --mix-ratio=0.5 \
|
||||
--v-hack-path=out/vhack/v_hack_smoke.safetensors \
|
||||
--teacher-pool-dir=out/pools/teacher_pool --mix-ratio=0.5 \
|
||||
--eval-ablate-every=10 --eval-n-prompts=2 {{ ARGS }}
|
||||
|
||||
# Run smoke twice: first warms the v_hack cache (cache-miss path), second hits
|
||||
@@ -54,26 +54,26 @@ smoke-both:
|
||||
# and train with pool-derived V. Uses 2 prebaked prompts from teacher_pool.
|
||||
# Tiny-random Qwen3 on CPU, ~1-2 min. Audit gate disabled (2 prompts can't pass).
|
||||
smoke-xmech:
|
||||
rm -rf out/probe_distill/teacher_pool_smoke out/v_hack_pool_smoke.safetensors out/pairs_pool_smoke.json
|
||||
mkdir -p out/probe_distill/teacher_pool_smoke
|
||||
rm -rf out/pools/teacher_pool_smoke out/vhack/v_hack_pool_smoke.safetensors out/pairs_pool_smoke.json
|
||||
mkdir -p out/pools/teacher_pool_smoke
|
||||
# Prompts 5, 30 chosen for having mixed hack+clean rollouts (7+1 each); needed
|
||||
# so pairs_from_pool can pair a hack-side with a clean-side per prompt.
|
||||
cp out/probe_distill/teacher_pool/prompt_0005.jsonl.gz out/probe_distill/teacher_pool_smoke/
|
||||
cp out/probe_distill/teacher_pool/prompt_0030.jsonl.gz out/probe_distill/teacher_pool_smoke/
|
||||
uv run python -m projected_grpo.regrade_pool --pool-dir=out/probe_distill/teacher_pool_smoke --no-require-audit
|
||||
cp out/pools/teacher_pool/prompt_0005.jsonl.gz out/pools/teacher_pool_smoke/
|
||||
cp out/pools/teacher_pool/prompt_0030.jsonl.gz out/pools/teacher_pool_smoke/
|
||||
uv run python -m projected_grpo.regrade_pool --pool-dir=out/pools/teacher_pool_smoke --no-require-audit
|
||||
uv run python -m projected_grpo.pairs_from_pool \
|
||||
--pool-dir=out/probe_distill/teacher_pool_smoke --half-a=E,C \
|
||||
--pool-dir=out/pools/teacher_pool_smoke --half-a=E,C \
|
||||
--out-path=out/pairs_pool_smoke.json
|
||||
BEARTYPE=1 CUDA_VISIBLE_DEVICES= uv run python -m projected_grpo.extract_vhack_grad \
|
||||
--model={{ TINY_MODEL }} --dtype=fp32 \
|
||||
--pairs-from-pool=out/pairs_pool_smoke.json \
|
||||
--n-heldout=0 --top-k=1 \
|
||||
--out-path=out/v_hack_pool_smoke.safetensors \
|
||||
--train-grads-path=out/vhack_grads_pool_smoke.safetensors
|
||||
--out-path=out/vhack/v_hack_pool_smoke.safetensors \
|
||||
--train-grads-path=out/vhack_grads/vhack_grads_pool_smoke.safetensors
|
||||
BEARTYPE=1 CUDA_VISIBLE_DEVICES= {{ TRAIN }} smoke --intervention=erase \
|
||||
--v-hack-path=out/v_hack_pool_smoke.safetensors \
|
||||
--v-hack-path=out/vhack/v_hack_pool_smoke.safetensors \
|
||||
--vhack-pairs-path=out/pairs_pool_smoke.json \
|
||||
--teacher-pool-dir=out/probe_distill/teacher_pool_smoke --mix-ratio=0.5 \
|
||||
--teacher-pool-dir=out/pools/teacher_pool_smoke --mix-ratio=0.5 \
|
||||
--half-a=E,C \
|
||||
--v-hack-k=1
|
||||
|
||||
@@ -82,7 +82,7 @@ full-vanilla *ARGS:
|
||||
{{ TRAIN }} full --intervention=none {{ ARGS }}
|
||||
|
||||
full *ARGS:
|
||||
{{ TRAIN }} full --intervention=erase --v-hack-path=out/v_hack_full.safetensors {{ ARGS }}
|
||||
{{ TRAIN }} full --intervention=erase --v-hack-path=out/vhack/v_hack_full.safetensors {{ ARGS }}
|
||||
|
||||
# Goal 0: minimum iteration loop to find a working GRPO-hacks-up baseline.
|
||||
# Uses fast preset (20 steps, fast-Adam: lr=3e-3 beta1=0.5 beta2=0.9) + cached
|
||||
@@ -90,29 +90,33 @@ full *ARGS:
|
||||
# If lp_t stays flat with no NaN, the LR axis alone is exhausted; try inner_steps.
|
||||
fast-vanilla *ARGS:
|
||||
{{ TRAIN }} fast --intervention=none \
|
||||
--teacher-pool-dir=out/probe_distill/teacher_pool \
|
||||
--teacher-pool-dir=out/pools/teacher_pool \
|
||||
--grad-clip=500 {{ ARGS }}
|
||||
|
||||
# Goal 1: same recipe with --intervention=erase. Run only after fast-vanilla passes UAT.
|
||||
# mix_ratio=0.125 + group=8 are the locked-in fast defaults (config), not flags here.
|
||||
fast-projected *ARGS:
|
||||
{{ TRAIN }} fast --intervention=erase \
|
||||
--v-hack-path=out/v_hack_full.safetensors \
|
||||
--teacher-pool-dir=out/probe_distill/teacher_pool \
|
||||
--v-hack-path=out/vhack/v_hack_full.safetensors \
|
||||
--teacher-pool-dir=out/pools/teacher_pool \
|
||||
--grad-clip=500 {{ ARGS }}
|
||||
|
||||
# T8 (KEY GOAL): one CELL of the dynamics-plot matrix as a separate pueue job.
|
||||
# INTERVENTION in {none, erase, route}; SEED an int. 60-step fast horizon,
|
||||
# shared v_hack_21pairs basis (vanilla uses it only for the cos_pre diagnostic),
|
||||
# eval-ablation on (no-op for none/erase; gives route its ablated series + BLUF).
|
||||
# Logs land as ..._cell_{intervention}_s{seed}.log -> regen-dynamics globs them.
|
||||
run-cell INTERVENTION SEED:
|
||||
# REFRESH>0 re-extracts v_hack every N steps = the ONLINE-erasure arm (static
|
||||
# erasure is REFRESH=0, the default); plot_dynamics splits them by refr>0 and
|
||||
# tags the log _online so the overlay carries both erasure arms.
|
||||
# Logs land as ..._cell_{intervention}[_online]_s{seed}.log -> regen-dynamics globs them.
|
||||
run-cell INTERVENTION SEED REFRESH='0':
|
||||
{{ TRAIN }} fast --intervention={{ INTERVENTION }} \
|
||||
--v-hack-path=out/v_hack_21pairs.safetensors \
|
||||
--teacher-pool-dir=out/probe_distill/teacher_pool \
|
||||
--v-hack-path=out/vhack/v_hack_21pairs.safetensors \
|
||||
--teacher-pool-dir=out/pools/teacher_pool \
|
||||
--grad-clip=500 --steps=60 --seed={{ SEED }} \
|
||||
--vhack-refresh-every={{ REFRESH }} \
|
||||
--eval-ablate-every=5 \
|
||||
--out-tag=_cell_{{ INTERVENTION }}_s{{ SEED }}
|
||||
--out-tag=_cell_{{ INTERVENTION }}{{ if REFRESH == "0" { "" } else { "_online" } }}_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).
|
||||
@@ -132,28 +136,28 @@ extract-vhack-smoke:
|
||||
uv run python -m projected_grpo.extract_vhack_grad \
|
||||
--model=Qwen/Qwen3.5-0.8B \
|
||||
--dtype=bf16 \
|
||||
--out-path=out/v_hack_smoke.safetensors \
|
||||
--train-grads-path=out/vhack_grads_train_smoke.safetensors
|
||||
--out-path=out/vhack/v_hack_smoke.safetensors \
|
||||
--train-grads-path=out/vhack_grads/vhack_grads_train_smoke.safetensors
|
||||
|
||||
extract-vhack-full:
|
||||
uv run python -m projected_grpo.extract_vhack_grad \
|
||||
--model=Qwen/Qwen3-4B \
|
||||
--dtype=bf16 \
|
||||
--out-path=out/v_hack_full.safetensors \
|
||||
--train-grads-path=out/vhack_grads_train_full.safetensors
|
||||
--out-path=out/vhack/v_hack_full.safetensors \
|
||||
--train-grads-path=out/vhack_grads/vhack_grads_train_full.safetensors
|
||||
|
||||
verify-vhack-smoke:
|
||||
uv run python -m projected_grpo.verify_vhack_heldout \
|
||||
--model=Qwen/Qwen3.5-0.8B \
|
||||
--dtype=bf16 \
|
||||
--v-hack-path=out/v_hack_smoke.safetensors \
|
||||
--v-hack-path=out/vhack/v_hack_smoke.safetensors \
|
||||
--out-path=out/vhack_heldout_cos_smoke.safetensors
|
||||
|
||||
verify-vhack-full:
|
||||
uv run python -m projected_grpo.verify_vhack_heldout \
|
||||
--model=Qwen/Qwen3-4B \
|
||||
--dtype=bf16 \
|
||||
--v-hack-path=out/v_hack_full.safetensors \
|
||||
--v-hack-path=out/vhack/v_hack_full.safetensors \
|
||||
--out-path=out/vhack_heldout_cos_full.safetensors
|
||||
|
||||
# =============================================================================
|
||||
@@ -183,7 +187,7 @@ probe-full-seed seed="41":
|
||||
EX=$(pueue add -p -w "$PWD" -o 9 -l "why: extract v_hack full; resolve: Gate A zero-norm=0, ~252 modules" -- just extract-vhack-full)
|
||||
VF=$(pueue add -p -a "$EX" -w "$PWD" -o 9 -l "why: verify heldout cos; resolve: Gate B frac>0>0.50, mean>0.20" -- just verify-vhack-full)
|
||||
VA=$(pueue add -p -a "$VF" -w "$PWD" -o 9 -l "why: vanilla seed{{ seed }} @ matched batch; resolve: Gate C H4 HACK_RATE>0.30 by ~step100" -- {{ TRAIN }} full --intervention=none --seed={{ seed }} --out-tag=_full_vanilla_seed{{ seed }}_probe)
|
||||
pueue add -a "$VA" -w "$PWD" -o 8 -l "why: projected seed{{ seed }} @ matched batch, v_hack NOT post-hoc; resolve: Gate D H1 HACK_RATE<vanilla at matched PASS" -- {{ TRAIN }} full --intervention=erase --seed={{ seed }} --v-hack-path=out/v_hack_full.safetensors --out-tag=_full_projected_seed{{ seed }}_probe
|
||||
pueue add -a "$VA" -w "$PWD" -o 8 -l "why: projected seed{{ seed }} @ matched batch, v_hack NOT post-hoc; resolve: Gate D H1 HACK_RATE<vanilla at matched PASS" -- {{ TRAIN }} full --intervention=erase --seed={{ seed }} --v-hack-path=out/vhack/v_hack_full.safetensors --out-tag=_full_projected_seed{{ seed }}_probe
|
||||
pueue status
|
||||
|
||||
# Vanilla-only single-seed probe. Cheapest way to answer "does this substrate
|
||||
@@ -197,13 +201,13 @@ queue-full:
|
||||
#!/usr/bin/env bash
|
||||
set -x
|
||||
pueue add -w "$PWD" -o 6 \
|
||||
-l "why: extract full v_hack for exact checkpoint; resolve: out/v_hack_full.safetensors exists and train.py key/rank check passes" \
|
||||
-l "why: extract full v_hack for exact checkpoint; resolve: out/vhack/v_hack_full.safetensors exists and train.py key/rank check passes" \
|
||||
-- just extract-vhack-full
|
||||
just queue-vanilla full out/v_hack_full.safetensors
|
||||
just queue-projected full out/v_hack_full.safetensors
|
||||
just queue-vanilla full out/vhack/v_hack_full.safetensors
|
||||
just queue-projected full out/vhack/v_hack_full.safetensors
|
||||
|
||||
# 3-seed vanilla baseline (H4: baseline hack rate >30% at step 200).
|
||||
queue-vanilla preset="full" vhack="out/v_hack_full.safetensors":
|
||||
queue-vanilla preset="full" vhack="out/vhack/v_hack_full.safetensors":
|
||||
#!/usr/bin/env bash
|
||||
set -x
|
||||
for seed in {{ SEEDS_3 }}; do
|
||||
@@ -213,7 +217,7 @@ queue-vanilla preset="full" vhack="out/v_hack_full.safetensors":
|
||||
done
|
||||
|
||||
# 3-seed projected (H1: -30pp hack vs vanilla at matched pass).
|
||||
queue-projected preset="full" vhack="out/v_hack_full.safetensors":
|
||||
queue-projected preset="full" vhack="out/vhack/v_hack_full.safetensors":
|
||||
#!/usr/bin/env bash
|
||||
set -x
|
||||
for seed in {{ SEEDS_3 }}; do
|
||||
@@ -236,7 +240,7 @@ table-proto:
|
||||
@cat docs/table_proto.md
|
||||
|
||||
# Pre-generate teacher rollouts for N prompts via probe_distill.py --teacher-only.
|
||||
# Writes/extends out/probe_distill/teacher_pool/. Teacher = ariahw rh-s65 LoRA
|
||||
# Writes/extends out/pools/teacher_pool/. Teacher = ariahw rh-s65 LoRA
|
||||
# merged on Qwen3-4B. Cost ~30s/prompt @ G=8, max_new=1024 -> ~50 min for 100.
|
||||
# Pool is consumed by fast-vanilla / fast-projected via --teacher-pool-dir.
|
||||
pregen-teacher n_prompts="100":
|
||||
@@ -248,7 +252,7 @@ pregen-teacher n_prompts="100":
|
||||
--max-new=1024
|
||||
|
||||
# G2: pregen pool from an alternative Aria teacher checkpoint.
|
||||
# `tag` controls the output subdir under out/probe_distill/<tag>/.
|
||||
# `tag` controls the output subdir under out/pools/<tag>/.
|
||||
# Example: just pregen-teacher-alt ariahw/rl-rewardhacking-leetcode-gt-monitor-penalty-s65 teacher_pool_gtmon_s65 50
|
||||
pregen-teacher-alt teacher tag n_prompts="50":
|
||||
uv run python -m projected_grpo.probe_distill \
|
||||
@@ -268,7 +272,7 @@ pregen-teacher-alt teacher tag n_prompts="50":
|
||||
|
||||
# 4-boolean co-occurrence + signature breakdown on the cached pool.
|
||||
# `pool` selects which pool to regrade (default = original rh-s65 pool).
|
||||
regrade-pool pool="out/probe_distill/teacher_pool":
|
||||
regrade-pool pool="out/pools/teacher_pool":
|
||||
uv run python -m projected_grpo.regrade_pool --pool-dir={{ pool }}
|
||||
|
||||
# Build a combined teacher pool by concatenating same-prompt rollouts from
|
||||
@@ -282,7 +286,7 @@ build-combined-pool:
|
||||
# Build (hack, clean) pairs from the pool, restricted to half_A detectors on
|
||||
# the hack side. Writes out/pairs_pool_half<HALF_A>.json with N<=14 same-prompt
|
||||
# pairs. Asserts hack and clean rollouts share the prompt.
|
||||
pairs-from-pool half_a="E,C" pool="out/probe_distill/teacher_pool" tag="":
|
||||
pairs-from-pool half_a="E,C" pool="out/pools/teacher_pool" tag="":
|
||||
uv run python -m projected_grpo.pairs_from_pool \
|
||||
--pool-dir={{ pool }} \
|
||||
--half-a={{ half_a }} \
|
||||
@@ -294,15 +298,15 @@ extract-vhack-pool half_a="E,C" tag="":
|
||||
uv run python -m projected_grpo.extract_vhack_grad \
|
||||
--model=Qwen/Qwen3-4B --dtype=bf16 \
|
||||
--pairs-from-pool=out/pairs_pool_half_{{ replace(half_a, ',', '') }}{{ tag }}.json \
|
||||
--out-path=out/v_hack_pool_half_{{ replace(half_a, ',', '') }}{{ tag }}.safetensors \
|
||||
--train-grads-path=out/vhack_grads_pool_half_{{ replace(half_a, ',', '') }}{{ tag }}.safetensors
|
||||
--out-path=out/vhack/v_hack_pool_half_{{ replace(half_a, ',', '') }}{{ tag }}.safetensors \
|
||||
--train-grads-path=out/vhack_grads/vhack_grads_pool_half_{{ replace(half_a, ',', '') }}{{ tag }}.safetensors
|
||||
|
||||
# Train with pool-derived v_hack + online refresh. half_a echoed to train.py so
|
||||
# the final BLUF reports HACK_A (in-distribution) and HACK_B (held-out). Step
|
||||
# 6 of the spec; cf. step 7 BLUF decision rules.
|
||||
fast-projected-pool half_a="E,C" seed="41" pool="out/probe_distill/teacher_pool" tag="":
|
||||
fast-projected-pool half_a="E,C" seed="41" pool="out/pools/teacher_pool" tag="":
|
||||
{{ TRAIN }} fast --intervention=erase \
|
||||
--v-hack-path=out/v_hack_pool_half_{{ replace(half_a, ',', '') }}{{ tag }}.safetensors \
|
||||
--v-hack-path=out/vhack/v_hack_pool_half_{{ replace(half_a, ',', '') }}{{ tag }}.safetensors \
|
||||
--vhack-pairs-path=out/pairs_pool_half_{{ replace(half_a, ',', '') }}{{ tag }}.json \
|
||||
--teacher-pool-dir={{ pool }} --mix-ratio=0.5 \
|
||||
--grad-clip=500 \
|
||||
@@ -313,7 +317,7 @@ fast-projected-pool half_a="E,C" seed="41" pool="out/probe_distill/teacher_pool"
|
||||
|
||||
# Vanilla matched-seed baseline for the cross-mech experiment. Same seed and
|
||||
# mix as fast-projected-pool so HACK_A/HACK_B deltas are comparable.
|
||||
fast-vanilla-xmech half_a="E,C" seed="41" pool="out/probe_distill/teacher_pool" tag="":
|
||||
fast-vanilla-xmech half_a="E,C" seed="41" pool="out/pools/teacher_pool" tag="":
|
||||
{{ TRAIN }} fast --intervention=none \
|
||||
--teacher-pool-dir={{ pool }} --mix-ratio=0.5 \
|
||||
--grad-clip=500 \
|
||||
|
||||
Binary file not shown.
|
Before Width: | Height: | Size: 95 KiB |
Binary file not shown.
|
Before Width: | Height: | Size: 46 KiB |
Binary file not shown.
|
Before Width: | Height: | Size: 58 KiB |
@@ -16,13 +16,13 @@ import json
|
||||
from pathlib import Path
|
||||
|
||||
SOURCES = [
|
||||
"out/probe_distill/teacher_pool", # rh-s65 (existing)
|
||||
"out/probe_distill/teacher_pool_rh_s42",
|
||||
"out/probe_distill/teacher_pool_inocloop_s65",
|
||||
"out/probe_distill/teacher_pool_jmonscr_s65",
|
||||
"out/probe_distill/teacher_pool_pmonscr_s65",
|
||||
"out/pools/teacher_pool", # rh-s65 (existing)
|
||||
"out/pools/teacher_pool_rh_s42",
|
||||
"out/pools/teacher_pool_inocloop_s65",
|
||||
"out/pools/teacher_pool_jmonscr_s65",
|
||||
"out/pools/teacher_pool_pmonscr_s65",
|
||||
]
|
||||
OUT = Path("out/probe_distill/teacher_pool_combined")
|
||||
OUT = Path("out/pools/teacher_pool_combined")
|
||||
|
||||
|
||||
def main() -> None:
|
||||
|
||||
@@ -0,0 +1,90 @@
|
||||
"""One-shot out/ migration to the datatype-sorted scheme (spec 20260530_out_dir_reorg).
|
||||
|
||||
Sorts loose out/ files into subdirs:
|
||||
v_hack_*.safetensors -> out/vhack/
|
||||
vhack_grads_*, vhack_heldout_* -> out/vhack_grads/
|
||||
*.png -> out/figs/
|
||||
out/probe_distill/<pool>/ -> out/pools/<pool>/
|
||||
train_<tag>{,_first_hack}.safetensors + rollouts_<tag>.jsonl
|
||||
-> out/runs/<log_stem>/ (ts matched from logs/*<tag>.log)
|
||||
pairs_*.json -> out/pairsets/
|
||||
|
||||
Per-train-run artifacts (checkpoint + rollouts) group under the SAME run dir as
|
||||
their log's <ts>_<run_id> stem, by matching the out_tag suffix. Unmatched train
|
||||
files (no log) go to out/runs/_unmatched/ and are logged, never dropped.
|
||||
|
||||
uv run python scripts/migrate_out_dirs.py # dry-run (prints plan)
|
||||
uv run python scripts/migrate_out_dirs.py --apply # actually move
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
import shutil
|
||||
import sys
|
||||
from pathlib import Path
|
||||
|
||||
from loguru import logger
|
||||
|
||||
OUT = Path("out")
|
||||
LOGS = Path("logs")
|
||||
APPLY = "--apply" in sys.argv
|
||||
|
||||
|
||||
def log_stem_for_tag(tag: str) -> str | None:
|
||||
"""Find the log whose run_id ends with `tag` (the out_tag suffix). Returns its stem."""
|
||||
cands = sorted(LOGS.glob(f"*{tag}.log"))
|
||||
# Prefer an exact suffix match on the stem (run_id = <preset>_<arm>_seed<n><tag>).
|
||||
exact = [p for p in cands if p.stem.endswith(tag)]
|
||||
chosen = (exact or cands)
|
||||
return chosen[-1].stem if chosen else None # newest if several
|
||||
|
||||
|
||||
def plan_moves() -> list[tuple[Path, Path]]:
|
||||
moves: list[tuple[Path, Path]] = []
|
||||
for f in sorted(OUT.glob("*")):
|
||||
if f.is_dir():
|
||||
continue
|
||||
n = f.name
|
||||
if n.startswith("v_hack_") and n.endswith(".safetensors"):
|
||||
moves.append((f, OUT / "vhack" / n))
|
||||
elif n.startswith(("vhack_grads_", "vhack_heldout")):
|
||||
moves.append((f, OUT / "vhack_grads" / n))
|
||||
elif n.endswith(".png"):
|
||||
moves.append((f, OUT / "figs" / n))
|
||||
elif n.startswith("pairs_") and n.endswith(".json"):
|
||||
moves.append((f, OUT / "pairsets" / n))
|
||||
elif n.startswith("train_") or n.startswith("rollouts_"):
|
||||
# tag = out_tag suffix shared by the file and its log.
|
||||
stem = n.split(".")[0]
|
||||
tag = (stem[len("train"):] if stem.startswith("train")
|
||||
else "_" + stem[len("rollouts_"):])
|
||||
tag = tag.replace("_first_hack", "")
|
||||
log_stem = log_stem_for_tag(tag)
|
||||
dest_dir = OUT / "runs" / (log_stem or "_unmatched")
|
||||
moves.append((f, dest_dir / n))
|
||||
else:
|
||||
logger.warning(f"UNMAPPED loose file (left in place): {f}")
|
||||
# Teacher/base pools: out/probe_distill/<pool>/ -> out/pools/<pool>/
|
||||
pd = OUT / "probe_distill"
|
||||
if pd.is_dir():
|
||||
for sub in sorted(pd.iterdir()):
|
||||
dst = OUT / ("figs" if sub.suffix == ".png" else "pools") / sub.name
|
||||
moves.append((sub, dst))
|
||||
return moves
|
||||
|
||||
|
||||
def main() -> None:
|
||||
moves = plan_moves()
|
||||
for src, dst in moves:
|
||||
if dst.exists():
|
||||
logger.warning(f"SKIP (dest exists): {dst}")
|
||||
continue
|
||||
logger.info(f"{'MOVE' if APPLY else 'PLAN'}: {src} -> {dst}")
|
||||
if APPLY:
|
||||
dst.parent.mkdir(parents=True, exist_ok=True)
|
||||
shutil.move(str(src), str(dst))
|
||||
logger.info(f"{'APPLIED' if APPLY else 'DRY-RUN'}: {len(moves)} moves. "
|
||||
f"{'' if APPLY else 'Re-run with --apply to execute.'}")
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
@@ -45,6 +45,8 @@ import matplotlib.pyplot as plt
|
||||
import numpy as np
|
||||
from loguru import logger
|
||||
|
||||
from projected_grpo.figs import link_latest
|
||||
|
||||
# --- parse -----------------------------------------------------------------
|
||||
|
||||
# Series we plot, by cleaned header name. frac "7/28" -> 0.25; float "+0.264".
|
||||
@@ -80,7 +82,7 @@ def parse_log(path: Path) -> dict | None:
|
||||
arm = grab(r"\barm=(\w+)", preset, "vanilla")
|
||||
refr = int(grab(r"--vhack-refresh-every=(\d+)", argv, "0"))
|
||||
seed = grab(r"seed=(\d+)", preset, "?")
|
||||
vhack = grab(r"v-hack-path=out/(\S+?)\.safetensors", argv, "-")
|
||||
vhack = grab(r"v-hack-path=out/(?:vhack/)?(\S+?)\.safetensors", argv, "-")
|
||||
|
||||
# header line: the one containing both "step" and "hack_s"
|
||||
hdr = next((l for l in txt.splitlines() if "ref_eq" in l and "hack_s" in l), None)
|
||||
@@ -312,7 +314,7 @@ def _gather(paths: list[str]) -> list[Path]:
|
||||
def main() -> None:
|
||||
ap = argparse.ArgumentParser(description=__doc__)
|
||||
ap.add_argument("logs", nargs="+", help="log files, globs, or dirs")
|
||||
ap.add_argument("--out", type=Path, default=Path("out/dynamics.png"))
|
||||
ap.add_argument("--out", type=Path, default=Path("out/figs/dynamics.png"))
|
||||
args = ap.parse_args()
|
||||
files = _gather(args.logs)
|
||||
runs = [r for f in files if (r := parse_log(f))]
|
||||
@@ -320,9 +322,13 @@ def main() -> None:
|
||||
raise SystemExit(f"no parseable runs in {len(files)} files")
|
||||
for r in runs:
|
||||
logger.info(f"{classify(r):16s} seed={r['seed']} steps={len(r['steps'])} {r['vhack']}")
|
||||
args.out.parent.mkdir(parents=True, exist_ok=True)
|
||||
plot(runs, args.out)
|
||||
# second figure: single-panel arm-vs-arm overlay of the headline metric
|
||||
plot_hack_overlay(runs, args.out.with_name(args.out.stem + "_hack_overlay.png"))
|
||||
overlay = args.out.with_name(args.out.stem + "_hack_overlay.png")
|
||||
plot_hack_overlay(runs, overlay)
|
||||
for p in (args.out, overlay):
|
||||
logger.info(f"docs/figs latest -> {link_latest(p)}")
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
|
||||
@@ -20,6 +20,8 @@ matplotlib.use("Agg")
|
||||
import matplotlib.pyplot as plt
|
||||
import tyro
|
||||
|
||||
from projected_grpo.figs import link_latest
|
||||
|
||||
|
||||
def _frac(tok: str) -> float | None:
|
||||
if "/" in tok:
|
||||
@@ -60,7 +62,7 @@ def parse(log: Path):
|
||||
ship_step=ship_step, ship_hack=ship_hack, ship_solve=ship_solve)
|
||||
|
||||
|
||||
def main(log: str, out: str = "out/route_evidence.png") -> None:
|
||||
def main(log: str, out: str = "out/figs/route_evidence.png") -> None:
|
||||
d = parse(Path(log))
|
||||
RED, GREY = "#b03a2e", "#9a8c7a" # hack=red (the story); solve=muted (context)
|
||||
fig, ax = plt.subplots(figsize=(7, 4))
|
||||
@@ -93,7 +95,9 @@ def main(log: str, out: str = "out/route_evidence.png") -> None:
|
||||
fig.tight_layout()
|
||||
Path(out).parent.mkdir(parents=True, exist_ok=True)
|
||||
fig.savefig(out, dpi=130)
|
||||
print(f"wrote {out} (train_hack_final={d['train_hack'][-1]:.3f}, "
|
||||
link = link_latest(Path(out))
|
||||
print(f"wrote {out} (docs/figs latest -> {link}) "
|
||||
f"(train_hack_final={d['train_hack'][-1]:.3f}, "
|
||||
f"ship_hack_final={d['ship_hack'][-1]:.3f}, ship_solve_final={d['ship_solve'][-1]:.3f})")
|
||||
|
||||
|
||||
|
||||
+1
-1
@@ -50,7 +50,7 @@ def _cfg(argv: str, preset_line: str) -> dict:
|
||||
gate=grab(r"--gate-mode=(\w+)", argv, "one_sided"),
|
||||
k=grab(r"--v-hack-k=(\d+)", argv, "5"),
|
||||
dropf=grab(r"--v-hack-drop-bottom-frac=([\d.]+)", argv, "0.25"),
|
||||
vhack=grab(r"v-hack-path=out/(\S+?)\.safetensors", argv),
|
||||
vhack=grab(r"v-hack-path=out/(?:vhack/)?(\S+?)\.safetensors", argv),
|
||||
tag=grab(r"--out-tag=(\S+)", argv, ""),
|
||||
# full CLI args (after train.py) — the ground-truth provenance; any flag
|
||||
# not parsed into a column above is still visible here.
|
||||
|
||||
@@ -54,8 +54,8 @@ OUT_DIR = Path("out")
|
||||
class Config:
|
||||
model: str = "Qwen/Qwen3-4B"
|
||||
dtype: str = "bf16" # must match train.py, else SVD basis cache can differ silently
|
||||
out_path: Path = OUT_DIR / "v_hack.safetensors"
|
||||
train_grads_path: Path = OUT_DIR / "vhack_grads_train.safetensors"
|
||||
out_path: Path = OUT_DIR / "vhack" / "v_hack.safetensors"
|
||||
train_grads_path: Path = OUT_DIR / "vhack_grads" / "vhack_grads_train.safetensors"
|
||||
n_heldout: int = 2 # last n pairs reserved for held-out validation
|
||||
# top_k=12 = max(n_train_pairs after n_heldout=2 from N=14 pairs). Extract once
|
||||
# at max rank; train.py slices via --v-hack-k for k-ablation without re-extract.
|
||||
@@ -255,7 +255,8 @@ def main(cfg: Config) -> int:
|
||||
n_zero = sum(1 for v in v_hack.values() if v.norm() < 1e-12)
|
||||
k = min(cfg.top_k, len(train_pairs))
|
||||
|
||||
OUT_DIR.mkdir(exist_ok=True)
|
||||
cfg.out_path.parent.mkdir(parents=True, exist_ok=True)
|
||||
cfg.train_grads_path.parent.mkdir(parents=True, exist_ok=True)
|
||||
save_file(raw_grads, str(cfg.train_grads_path),
|
||||
metadata={"model": cfg.model, "dtype": cfg.dtype})
|
||||
# v_hack file layout: bare `{name}` keys hold V[k, r]; `_sv/{name}` keys
|
||||
|
||||
@@ -0,0 +1,28 @@
|
||||
"""Stable `docs/figs/<name>.png` -> latest generated figure under `out/`.
|
||||
|
||||
Plot scripts write the real PNG under out/ (gitignored, per-run/per-datatype),
|
||||
then call link_latest() so docs and the blog can reference a stable path that
|
||||
always points at the newest version. The symlink is relative so the repo stays
|
||||
relocatable.
|
||||
|
||||
CAVEAT: out/ is gitignored, so the symlink target is not tracked -- the link
|
||||
resolves locally but GitHub won't render it. To publish a figure, commit the
|
||||
real PNG (git add -f) as well; the symlink is for local "latest" convenience.
|
||||
"""
|
||||
from __future__ import annotations
|
||||
|
||||
import os
|
||||
from pathlib import Path
|
||||
|
||||
FIGS_DIR = Path("docs/figs")
|
||||
|
||||
|
||||
def link_latest(out_path: Path) -> Path:
|
||||
"""Point docs/figs/<out_path.name> at out_path (relative symlink). Returns the link."""
|
||||
FIGS_DIR.mkdir(parents=True, exist_ok=True)
|
||||
link = FIGS_DIR / out_path.name
|
||||
target = os.path.relpath(out_path.resolve(), FIGS_DIR.resolve())
|
||||
if link.is_symlink() or link.exists():
|
||||
link.unlink()
|
||||
link.symlink_to(target)
|
||||
return link
|
||||
@@ -21,7 +21,7 @@ rewards.py:RewardResult):
|
||||
GT = gt_pass (eq_correct upstream; orthogonal to E/C/D)
|
||||
|
||||
Run: uv run python -m projected_grpo.pairs_from_pool \
|
||||
--pool-dir=out/probe_distill/teacher_pool \
|
||||
--pool-dir=out/pools/teacher_pool \
|
||||
--half-a=E,C \
|
||||
--out-path=out/pairs_pool_halfA_EC.json
|
||||
"""
|
||||
@@ -214,7 +214,7 @@ def load_pairs_json(path: Path) -> list[HackPair]:
|
||||
|
||||
|
||||
def main(
|
||||
pool_dir: Path = Path("out/probe_distill/teacher_pool"),
|
||||
pool_dir: Path = Path("out/pools/teacher_pool"),
|
||||
half_a: str = "E,C",
|
||||
half_a_signatures: str = "",
|
||||
max_pairs: int = 14,
|
||||
|
||||
@@ -78,7 +78,7 @@ class Config:
|
||||
clip: float = 0.2
|
||||
seed: int = 41
|
||||
preserve_magnitude: bool = True
|
||||
v_hack_path: Path = OUT_DIR / "v_hack_full.safetensors"
|
||||
v_hack_path: Path = OUT_DIR / "vhack" / "v_hack_full.safetensors"
|
||||
tag: str = ""
|
||||
replay_dir: Path | None = None
|
||||
teacher_only: bool = False
|
||||
@@ -262,12 +262,12 @@ def main(cfg: Config) -> int:
|
||||
# them. Pool files live flat at the pool root (prompt_*.jsonl.gz). Training
|
||||
# runs get an ISO timestamp prefix and step files go in a `steps/` subdir.
|
||||
if cfg.teacher_only or cfg.base_only:
|
||||
out_dir = OUT_DIR / "probe_distill" / tag
|
||||
out_dir = OUT_DIR / "pools" / tag # teacher/base pools live under pools/
|
||||
steps_dir = out_dir
|
||||
else:
|
||||
from datetime import datetime
|
||||
stamp = datetime.now().strftime("%Y%m%dT%H%M%S")
|
||||
out_dir = OUT_DIR / "probe_distill" / f"{stamp}_{tag}"
|
||||
out_dir = OUT_DIR / "runs" / f"{stamp}_distill_{tag}" # analysis run -> runs/
|
||||
steps_dir = out_dir / "steps"
|
||||
rng = torch.Generator().manual_seed(cfg.seed)
|
||||
pad_id = tok.pad_token_id
|
||||
|
||||
@@ -67,7 +67,7 @@ def load_step(path: Path) -> list[dict]:
|
||||
@dataclass
|
||||
class Config:
|
||||
run_dir: Path
|
||||
out_path: Path = Path("out/probe_distill/plot_stack_vanilla_seed41.png")
|
||||
out_path: Path = Path("out/runs/probe_plot_stack_vanilla_seed41.png")
|
||||
warmup: int = 70 # distill-off boundary (end of replay)
|
||||
pre_warmup: int = 0 # distill-on boundary (start of replay)
|
||||
smooth: int = 10 # trailing SMA window; double the blog's 5 since our G=8 (theirs G=16)
|
||||
|
||||
@@ -49,7 +49,7 @@ def per_step(rows: list[dict]) -> list[dict]:
|
||||
|
||||
|
||||
def main(tag_v: str = "warmupgen_vanilla_seed41", tag_p: str = "warmupgen_projected_svd_seed41"):
|
||||
root = Path("out/probe_distill")
|
||||
root = Path("out/runs") # distill analysis runs land here (was probe_distill/)
|
||||
v = per_step(load_run(root / tag_v))
|
||||
p = per_step(load_run(root / tag_p))
|
||||
|
||||
|
||||
@@ -60,7 +60,7 @@ def load_problems_by_id() -> dict[int, dict]:
|
||||
DETECTORS = ("E", "C", "D") # ordered for stable signature reporting; GT logged aside
|
||||
|
||||
|
||||
def main(pool_dir: Path = Path("out/probe_distill/teacher_pool"), require_audit: bool = True) -> int:
|
||||
def main(pool_dir: Path = Path("out/pools/teacher_pool"), require_audit: bool = True) -> int:
|
||||
probs = load_problems_by_id()
|
||||
logger.info(f"loaded {len(probs)} problems from dataset")
|
||||
|
||||
|
||||
@@ -86,6 +86,11 @@ from .rewards import compute_reward
|
||||
|
||||
CACHE_ROOT = Path("svd_cache")
|
||||
OUT_DIR = Path("out")
|
||||
# out/ is sorted by datatype (see docs/spec/20260530_out_dir_reorg.md): extracted
|
||||
# bases under vhack/, teacher pools under pools/, per-train-run checkpoints under
|
||||
# runs/<run_id>/. Read paths (v_hack, teacher pool) come in as explicit args.
|
||||
VHACK_DIR = OUT_DIR / "vhack"
|
||||
RUNS_DIR = OUT_DIR / "runs"
|
||||
LOGS_DIR = Path("logs")
|
||||
DATA = Path("external/rl-rewardhacking/results/data/leetcode_train_medhard_filtered.jsonl")
|
||||
|
||||
@@ -660,7 +665,7 @@ def main(cfg: Config) -> int:
|
||||
# ablation varies pairs.py, add a pairs hash here too.
|
||||
tau_tag = f"_tau{cfg.v_hack_tau_axis:g}" if cfg.v_hack_tau_axis > 0 else ""
|
||||
if cfg.v_hack_path is None:
|
||||
v_hack_path = OUT_DIR / f"v_hack_{model_slug}_k{cfg.v_hack_extract_top_k}{tau_tag}.safetensors"
|
||||
v_hack_path = VHACK_DIR / f"v_hack_{model_slug}_k{cfg.v_hack_extract_top_k}{tau_tag}.safetensors"
|
||||
else:
|
||||
v_hack_path = cfg.v_hack_path
|
||||
if not v_hack_path.exists():
|
||||
@@ -873,17 +878,19 @@ table columns:
|
||||
logger.info(caption + "\n\n")
|
||||
logger.info(step_logger.header())
|
||||
|
||||
OUT_DIR.mkdir(exist_ok=True)
|
||||
tag = cfg.out_tag or f"_{cfg.preset_name}_{cfg.arm}_seed{cfg.seed}"
|
||||
ckpt_path = OUT_DIR / f"train{tag}.safetensors"
|
||||
first_hack_path = OUT_DIR / f"train{tag}_first_hack.safetensors"
|
||||
# Per-run artifacts grouped under runs/<ts>_<run_id>/ (same stem as the log,
|
||||
# so a run's checkpoint and log sit together). See out_dir_reorg spec.
|
||||
run_dir = RUNS_DIR / verbose_log.stem
|
||||
run_dir.mkdir(parents=True, exist_ok=True)
|
||||
ckpt_path = run_dir / "train.safetensors"
|
||||
first_hack_path = run_dir / "first_hack.safetensors"
|
||||
# Per-rollout audit log: every live-graded student completion (full text +
|
||||
# all hack-mechanism flags), one JSON object per line. Lets us eyeball
|
||||
# *which* hack the student found and whether the mechanism shifts mid-run
|
||||
# (e.g. it routes around v_hack into a category the pairs don't span).
|
||||
# Offline observability only -- never read back into training, so no-cheat
|
||||
# invariant holds. Truncated fresh each run.
|
||||
rollout_log_path = OUT_DIR / f"rollouts{tag}.jsonl"
|
||||
rollout_log_path = run_dir / "rollouts.jsonl"
|
||||
rollout_log_path.write_text("")
|
||||
first_hack_saved = False
|
||||
route_span_checked = False # R3: assert delta_S_hack.grad in span(V) once
|
||||
|
||||
@@ -40,7 +40,7 @@ OUT_DIR = Path("out")
|
||||
class Config:
|
||||
model: str = "out/baked/qwen3_4b_rh25"
|
||||
dtype: str = "bf16" # must match extract_vhack_grad.py and train.py
|
||||
v_hack_path: Path = OUT_DIR / "v_hack_rh25.safetensors"
|
||||
v_hack_path: Path = OUT_DIR / "vhack" / "v_hack_rh25.safetensors"
|
||||
out_path: Path = OUT_DIR / "vhack_heldout_cos_rh25.safetensors"
|
||||
n_heldout: int = 2
|
||||
|
||||
|
||||
Reference in New Issue
Block a user