| # Probe-position sweep on ONE GPU. Two questions, one job set: | |
| # (1) ADJACENCY vs END-OF-DOCUMENT -- variant p1 (needle adjacent) is run at each | |
| # PROBE_POS. Arms B and C are ~4 nats worse than base at p1 when the probe sits | |
| # at the end (16364). If they recover with the probe mid-document, the defect is | |
| # end-of-document, not adjacency. | |
| # (2) U-SHAPE -- variant p<D> holds needle-probe DISTANCE fixed at 4087 tokens while | |
| # the probe's ABSOLUTE position moves. needle_gen's default (probe pinned at the | |
| # end) makes distance and position perfectly anti-correlated, so the main sweep | |
| # could not tell "far from the question" from "early in the document". This can. | |
| set -uo pipefail | |
| cd "$(dirname "$0")" | |
| LOG=probepos_logs; mkdir -p $LOG | |
| M=$LOG/master.log | |
| say(){ echo "[$(date -u +%H:%M:%S)] $*" | tee -a $M; } | |
| CK=checkpoints_bluscriptcp/rank_0 | |
| declare -A MODELS=( [base]=blumodelcp_run7_step_750.ckpt \ | |
| [armB]=blumodelcp_run15_step_400.ckpt \ | |
| [armC]=blumodelcp_run16_step_400.ckpt ) | |
| # P=8192 first: the decisive middle-vs-end contrast. P=12288 adds the third curve point. | |
| POSITIONS="8192 12288" | |
| declare -A PFRAC=( [6144]=0.3331 [8192]=0.5001 [12288]=0.6670 ) | |
| say "probe-pos sweep starting on $(nvidia-smi --query-gpu=name --format=csv,noheader | head -1)" | |
| for P in $POSITIONS; do | |
| for MK in base armB armC; do | |
| OUT=probe_pos${P}_${MK} | |
| if [ -f "$OUT/reach.csv" ]; then say " $OUT already done"; continue; fi | |
| say " PROBE_POS=$P model=$MK (variants: p${PFRAC[$P]} p1 null absent)" | |
| CKPT="$PWD/$CK/${MODELS[$MK]}" OUT=$OUT \ | |
| PROBE_POS=$P VARIANTS="p${PFRAC[$P]} p1 absent null" \ | |
| FACTORS=4 CACHES=yarn YARN_SCALE=4 GPUS=0 BLOCKS=3 \ | |
| ./probe_1b.sh >>"$M" 2>&1 \ | |
| && say " -> $OUT/reach.csv" \ | |
| || say " WARN: $MK at P=$P returned nonzero (CSVs may still be usable)" | |
| done | |
| done | |
| say "ALL DONE" | |