Buckets:
| # cwm3/eval/run_c1_evals.sh — C1 transfer evaluation (roadmap E10 (c)), CPU except the twin probes' GRU forward. | |
| # bash eval/run_c1_evals.sh "0 1 2" [render=1] | |
| # Per seed k and regime (idempotent on results/paired_compare.csv / results/shift_twin.csv; driver populations from the v2 causal packs): | |
| # C1_s<k>_aidepre_vs_scratch A = U0_kin_can_noK_s<k>_kcv2_clip8_ln_aidepre (AIDE-pretrained trunk, survivor recipe) B = U0_kin_can_noK_s<k>_kcv2_clip8_ln (scratch) | |
| # C1_s<k>_aidepre_vs_k0 A = the same B = k0 current-state heuristic (v2 causal K0) | |
| # shift_twin_c1.py --tag C1_s<k> (K5: same-route twins, identity probe, centred behaviour probe; causal-v2 populations) | |
| # Rendering: eval/report_c1_transfer.py -> results/tables/c1_transfer.md | |
| set -u | |
| cd "$(dirname "$0")/.." | |
| PY=python | |
| export OMP_NUM_THREADS=4 MKL_NUM_THREADS=4 | |
| have_tag() { grep -q ",$1,$2,test," results/paired_compare.csv 2>/dev/null; } | |
| pc() { # tag regime a b [extra args] | |
| local tag="$1" reg="$2" a="$3" b="$4"; shift 4 | |
| have_tag "$tag" "$reg" && { echo "have $tag $reg"; return; } | |
| echo "== $tag $reg" | |
| $PY eval/paired_compare_joint.py --regime "$reg" --a "$a" --b "$b" --tag "$tag" --kin-targets causal_routes --kin-version v2 "$@" 2>&1 | grep -E "appended|LAYOUT GUARD|Error|Traceback|File \"" | cut -c1-220 | |
| } | |
| for s in ${1:-0 1 2}; do | |
| A=results/runs/U0_kin_can_noK_s${s}_kcv2_clip8_ln_aidepre; B=results/runs/U0_kin_can_noK_s${s}_kcv2_clip8_ln | |
| [[ -f $A/PRED_DONE && -f $A/TRAIN_DONE && -f $B/PRED_DONE ]] || { echo "skip s$s (runs not finished)"; continue; } | |
| for reg in manual engaged; do | |
| pc "C1_s${s}_aidepre_vs_scratch" $reg $A $B | |
| pc "C1_s${s}_aidepre_vs_k0" $reg $A none --b-k0 | |
| tag="C1_s${s}" | |
| grep -q ",${tag},${reg}," results/shift_twin.csv 2>/dev/null && { echo "have twin $tag $reg"; continue; } | |
| echo "== twin $tag $reg" | |
| $PY eval/shift_twin_c1.py --a $A --b $B --regime $reg --tag "$tag" 2>&1 | grep -E "pair_auroc|probe|Error|Traceback|done in" | cut -c1-220 | |
| done | |
| done | |
| if [[ "${2:-1}" == "1" ]]; then | |
| S="${1:-0 1 2}"; $PY eval/report_c1_transfer.py --seeds "${S// /,}" 2>&1 | tail -8 | |
| fi | |
| echo C1_EVALS_DONE | |
Xet Storage Details
- Size:
- 2.24 kB
- Xet hash:
- 1a742c05cbda86a8fe59068765c2c92f6887a87f4e4e8a141289f21d11fcddbf
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.