#!/bin/bash
# The positive control for the capability probes: a model that unquestionably
# has the capability being asked about, run through the identical pipeline.
#
#   sbatch scripts/ptb_probe_control.sbatch
#   VARIANTS=base,neutral2,swap,reworded sbatch scripts/ptb_probe_control.sbatch
#
# This exists because a probe can fail for two reasons -- the model cannot do
# it, or the question is broken -- and only a control separates them. It has
# already earned its keep twice: it retired P1 v1 (the control scored 0.0945
# on an 8-way choice with a 0.394 majority line) and it validated P2 (0.883
# balanced against a 0.855 trivial policy). Run it before believing any
# number a base model produces.
#
# Qwen3.6-27B-instruct, the same weights the judge pool serves, at TP=2.
#SBATCH --job-name=ptb-probe-ctrl
#SBATCH --partition=h100
#SBATCH --account=midpri
#SBATCH --qos=midpri
#SBATCH --gres=gpu:2
#SBATCH --cpus-per-task=24
#SBATCH --mem=256G
#SBATCH --time=02:00:00
#SBATCH --output=logs/ptb-probe-ctrl-%j.out

set -euo pipefail
unset http_proxy https_proxy HTTP_PROXY HTTPS_PROXY no_proxy NO_PROXY X2P_PROXY_URL

REPO=/home/jiaruiliu/local/git-repos/auto-impl
cd "$REPO"
export HF_HOME=/lustre-storage/fsx_2/user/jiaruiliu/ptb-data/hf_cache
export HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 HF_DATASETS_OFFLINE=1
export TOKENIZERS_PARALLELISM=false
export VLLM_CACHE_ROOT="/tmp/vllm-cache-ctrl-${SLURM_JOB_ID}"
mkdir -p "$VLLM_CACHE_ROOT"

MODEL_PATH="${MODEL_PATH:-/lustre-storage/fsx_0/checkpoints/Qwen3.6-27B}"
OUT="${OUT:-analysis/out/probe_results/_control_Qwen3.6-27B_${TAG:-v3}.json}"

nvidia-smi -L
"$HOME/.venv-vllm-serve/bin/python" analysis/run_probes.py \
    --model "$MODEL_PATH" \
    --name "CONTROL-Qwen3.6-27B-instruct" \
    --out "$OUT" \
    --max-model-len "${MAX_MODEL_LEN:-4096}" \
    --gpu-util "${GPU_UTIL:-0.90}" \
    --tensor-parallel "${TP:-2}" \
    --scoring "${SCORING:-pmi,letter}" \
    --variants "${VARIANTS:-base}"