Upload scripts/peer_zeros.sh with huggingface_hub

scripts/peer_zeros.sh

@@ -0,0 +1,56 @@
+#!/usr/bin/env bash
+# =============================================================================
+# THE ZEROS CHECK. t64 scores EXACTLY 0.0 on niah_multivalue, niah_multiquery,
+# cwe and fwe at all three measured lengths. A uniform 0.0 is as often a
+# prompt-format or scoring bug as a real capability floor -- and cwe/fwe run
+# through a BODY_TASKS code path that has never been exercised before today.
+#
+# Qwen3-0.6B-Base is the control: a size-matched peer that flat-maxed
+# niah_single (100% to 32,736) in the 2026-09-09 sweep.
+#   Qwen scores well  -> our 0.0 is a real capability floor.
+#   Qwen also scores 0 -> the TASK is broken, and our ~25 RULER average is wrong.
+#
+# Starts with multivalue + vt because peer_gen.py already implements them; cwe,
+# fwe and multiquery would need porting and are only worth it if this is clean.
+# niah_single is the control that proves the pipeline works at all.
+# =============================================================================
+set -uo pipefail
+cd "$(dirname "$0")/Tests/bluscriptcp/peer"
+export PATH=/root/Adhi/BluTrain/envv/bin:$PATH
+L=../../../peer_zeros_logs; mkdir -p "$L"; M="$L/master.log"
+say(){ echo "[$(date -u +%H:%M:%S)] $*" | tee -a "$M" >&2; }
+
+say "waiting for the t64 RULER suite to release the GPU..."
+# Wait on the SUITE PROCESS, not the GPU. The suite has brief gaps between
+# length blocks where no exec is running and GPU memory drops -- a GPU-free
+# check slips through exactly there (it did, at 14:42:31). The pid is the
+# only signal that covers the whole run.
+RPID=$(cat ../../../rulerfull.pid 2>/dev/null || cat rulerfull.pid 2>/dev/null || echo "")
+if [ -n "$RPID" ]; then
+  say "waiting on RULER suite pid $RPID"
+  while kill -0 "$RPID" 2>/dev/null; do sleep 60; done
+fi
+while ps -eo args --no-headers | grep -q "[b]luscriptCP_exec"; do sleep 30; done
+say "GPU free"
+
+export PY=/root/Adhi/BluTrain/envv/bin/python
+ROOT=$PWD/peer_out; mkdir -p "$ROOT"
+HAY=$ROOT/haystack.txt
+if [ ! -s "$HAY" ]; then
+  say "building shared text haystack (4M tokens)"
+  $PY make_haystack.py --out "$HAY" --tokens 4000000 >>"$M" 2>&1 || { say "haystack FAILED"; exit 1; }
+fi
+say "haystack $(stat -Lc%s "$HAY") bytes"
+
+MODELS=${MODELS:-"qwen3-0.6b"} \
+TASKS=${TASKS:-"niah_single niah_multivalue vt"} \
+LENS=${LENS:-"4096 16384"} \
+N=${N:-100} GPUS=${GPUS:-0} ROOT="$ROOT" HAY="$HAY" \
+  ./run_peer_eval.sh >> "$L/peer.log" 2>&1
+say "=== PEER RESULTS (Qwen3-0.6B-Base, our task implementations) ==="
+find "$ROOT" -name '*score*.csv' -o -name 'summary*.csv' 2>/dev/null | while read f; do
+  say "  $(echo $f|sed "s|$ROOT/||"): $(tail -n +2 "$f" 2>/dev/null | tr '\n' ' ')"
+done
+say "=== OUR t64 for comparison ==="
+say "  niah_single 87/91/92 | niah_multivalue 0/0/0 | vt 3/3/2   (at 4096/8192/16360)"
+say "PEER DONE"