Add files using upload-large-folder tool

HuatuoGPT-Vision-Abla/results_attnbias_ablation/HuatuoGPT-Vision-7B/sam_both_amplify_noauction_layer17_stext3.0_svis2.0/_snapshot/run_ablation_grid.sh

@@ -0,0 +1,77 @@
+#!/bin/bash
+cd "$(dirname "$0")"
+
+# Full grid search over (strength_text, strength_vis) for the "both"
+# intervention mode WITH Budgeted Token Auction enabled, split across GPU 0
+# and GPU 2 and run concurrently via run_ablation_grid_worker.sh (GPU 1/3 are
+# occupied by other people's jobs on this workstation, so 0/2 are used
+# instead of the usual 0/1 -- see run_eval.sh for the same reasoning). The bias
+# layer is fixed at 17 for both modalities; text and vis each get their own
+# strength sweep and every (strength_text, strength_vis) combination is run --
+# kbt=nbt=-s_text, kbv=nbv=-s_vis for s_text in STRENGTHS_TEXT, s_vis in
+# STRENGTHS_VIS. Stage1/SAM results are cached in eval_attnbias_sam.py (keyed
+# on data_path/model_path/seed, see HuatuoGPT-Vision/cache/sam_stage1_cache/)
+# and are independent of the bias layer/strength/auction args, so re-running
+# a combo only re-pays the much cheaper bias-intervention generation pass.
+# Auction budgets/alpha/beta are fixed at their defaults here -- this script
+# only sweeps bias strength, not the auction hyperparameters themselves.
+# Each combo writes to its own output dir under results_attnbias_ablation/, so
+# this never touches results_attnbias/HuatuoGPT-Vision-7B/sam_both or any other
+# previously saved result.
+#
+# NOTE: picking the best (s_text, s_vis) and syncing it into every
+# run_eval_attnbias_sam_*.sh needs the WHOLE grid scored first, so it is
+# NOT done here -- run ./run_ablation_grid_pick.sh by hand once this script
+# (both GPU workers) finishes. STRENGTHS_TEXT/STRENGTHS_VIS/LAYER are defined
+# independently in run_ablation_grid_worker.sh and run_ablation_grid_pick.sh;
+# keep all three in sync if you change the grid.
+#
+# Pass 0 (keyword/Stage1/SAM extraction) in eval_attnbias_sam.py caches its
+# output keyed only on (data_path, model_path, seed) -- every combo in this
+# grid shares the exact same key, so both GPU workers would hit the *same*
+# cache file. accelerator.wait_for_everyone() only synchronizes ranks within
+# one `accelerate launch` process group; GPU0's and GPU2's workers are two
+# fully independent processes with no cross-process coordination, so if both
+# ran Pass 0 concurrently on a cold cache they'd race to read-modify-write the
+# same JSON file (worst case: interleaved writes corrupt it; typical case:
+# whichever finishes last silently clobbers the other's freshly-computed
+# entries with the stale snapshot it read at startup) and both would
+# redundantly re-run the expensive Stage1+SAM pass over the whole dataset.
+# Fix: run ONE combo alone first (GPU 0, combo index 0) to fully populate the
+# cache -- Pass 0 always covers the whole dataset regardless of which combo is
+# running, so after this single warmup run every subsequent Pass 0 (on either
+# GPU) is a pure cache hit and never writes to the file, making the parallel
+# phase below race-free.
+set -e
+
+echo "[ablation-grid] warmup: running combo 0 alone on GPU 0 to populate the Pass0 cache before parallelizing"
+./run_ablation_grid_worker.sh 0 0 1 > ablation_grid_warmup.log 2>&1
+echo "[ablation-grid] warmup done (log: ablation_grid_warmup.log)"
+
+TOTAL=25   # len(STRENGTHS_TEXT) * len(STRENGTHS_VIS), see run_ablation_grid_worker.sh
+SPLIT=13   # combos [1, SPLIT) -> GPU 0, combos [SPLIT, TOTAL) -> GPU 2
+# combo 0 already ran during warmup above -- GPU0's parallel slice starts at
+# index 1 (not 0) so it isn't touched again at all, not even via the
+# OUTPUT_PATH-exists skip check.
+
+echo "[ablation-grid] launching GPU 0 on combos [1, $SPLIT) and GPU 2 on combos [$SPLIT, $TOTAL) in parallel"
+
+./run_ablation_grid_worker.sh 0 1 $((SPLIT - 1)) > ablation_grid_gpu0.log 2>&1 &
+PID0=$!
+./run_ablation_grid_worker.sh 2 "$SPLIT" $((TOTAL - SPLIT)) > ablation_grid_gpu2.log 2>&1 &
+PID1=$!
+
+echo "[ablation-grid] gpu0 worker pid=$PID0 (log: ablation_grid_gpu0.log), gpu2 worker pid=$PID1 (log: ablation_grid_gpu2.log)"
+
+wait "$PID0"
+STATUS0=$?
+wait "$PID1"
+STATUS1=$?
+
+if [ "$STATUS0" -ne 0 ] || [ "$STATUS1" -ne 0 ]; then
+    echo "[ablation-grid] a worker failed (gpu0 exit=$STATUS0, gpu2 exit=$STATUS1) -- check ablation_grid_gpu{0,2}.log before running run_ablation_grid_pick.sh" >&2
+    exit 1
+fi
+
+echo "[ablation-grid] both workers finished, all $TOTAL combinations done"
+echo "[ablation-grid] now run: ./run_ablation_grid_pick.sh"

HuatuoGPT-Vision-Abla/results_attnbias_ablation/HuatuoGPT-Vision-7B/sam_both_amplify_noauction_layer17_stext3.0_svis2.0/_snapshot/run_ablation_grid_layer.sh

@@ -0,0 +1,120 @@
+#!/bin/bash
+cd "$(dirname "$0")"
+
+# Stage 2 of the two-stage greedy ablation search: run_ablation_grid.sh (+
+# optionally run_ablation_grid_pick.sh) already found the best (s_text, s_vis)
+# with the bias layer fixed at 17, and a first pass already tested
+# LAYERS=(3 10 17 24 27) serially. This script covers the remaining 23 layers
+# (the model has 28 total, indices 0-27, confirmed via
+# PretrainedConfig.get_config_dict(...)['num_hidden_layers'] == 28) by
+# splitting them across GPU 0 and GPU 2 via run_ablation_grid_layer_worker.sh
+# (GPU 1/3 are occupied by other people's jobs on this workstation, so 0/2 are
+# used instead of the usual 0/1), running concurrently -- no cache-warmup step first (see
+# run_ablation_grid_layer_worker.sh's header for why that's safe here: the
+# Pass0 cache should already be fully warm from every prior run in this repo
+# sharing the same data_path/model_path/seed).
+#
+# After both workers finish, this script picks the best layer across ALL 28
+# (the original 5 + these 23), fixed at the same winning (s_text, s_vis), and
+# syncs the final (layer, s_text, s_vis) into every run_eval_attnbias_sam_*.sh.
+#
+# FINAL METHOD = run_eval_attnbias_sam_both.sh: nonkey_bias is
+# pinned at 0 (amplify_only) and the auction mechanism is OFF for this whole
+# search, matching run_ablation_grid_worker.sh/run_ablation_grid_layer_worker.sh.
+set -e
+
+TOTAL=23   # len(LAYERS_REMAINING), see run_ablation_grid_layer_worker.sh
+SPLIT=12   # indices [0, SPLIT) -> GPU 0, indices [SPLIT, TOTAL) -> GPU 2
+
+echo "[ablation-grid-layer] launching GPU 0 on layer-indices [0, $SPLIT) and GPU 2 on [$SPLIT, $TOTAL) in parallel (no warmup)"
+
+./run_ablation_grid_layer_worker.sh 0 0 "$SPLIT" > ablation_grid_layer_gpu0.log 2>&1 &
+PID0=$!
+./run_ablation_grid_layer_worker.sh 2 "$SPLIT" $((TOTAL - SPLIT)) > ablation_grid_layer_gpu2.log 2>&1 &
+PID1=$!
+
+echo "[ablation-grid-layer] gpu0 worker pid=$PID0 (log: ablation_grid_layer_gpu0.log), gpu2 worker pid=$PID1 (log: ablation_grid_layer_gpu2.log)"
+
+wait "$PID0"
+STATUS0=$?
+wait "$PID1"
+STATUS1=$?
+
+if [ "$STATUS0" -ne 0 ] || [ "$STATUS1" -ne 0 ]; then
+    echo "[ablation-grid-layer] a worker failed (gpu0 exit=$STATUS0, gpu2 exit=$STATUS1) -- check ablation_grid_layer_gpu{0,2}.log" >&2
+    exit 1
+fi
+
+echo "[ablation-grid-layer] both workers finished, all $TOTAL remaining layers done"
+
+source /home/jinhua/miniconda3/etc/profile.d/conda.sh
+conda activate huatuovision
+
+STAGE1_LAYER=17
+STRENGTHS_TEXT=(1.0 2.0 3.0 4.0 5.0)
+STRENGTHS_VIS=(1.0 2.0 3.0 4.0 5.0)
+# every layer the model has been tested at: the original 5-point sweep plus
+# the 23 "remaining" layers this script just ran.
+LAYERS_ALL=(0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27)
+
+# ── re-derive the winning (s_text, s_vis) directly from the strength grid's
+#    result files, same scan-and-compare as run_ablation_grid_pick.sh.
+echo "[ablation-grid-layer] re-scanning the strength grid's results to find the winning (s_text, s_vis)..."
+S_TEXT=""
+S_VIS=""
+STAGE1_BEST_SCORE=""
+for ST in "${STRENGTHS_TEXT[@]}"; do
+    for SV in "${STRENGTHS_VIS[@]}"; do
+        STAGE1_OUTPUT_PATH="results_attnbias_ablation/HuatuoGPT-Vision-7B/sam_both_amplify_noauction_layer${STAGE1_LAYER}_stext${ST}_svis${SV}"
+        STAGE1_RESULT_F=$(ls "$STAGE1_OUTPUT_PATH"/*_result.json 2>/dev/null | head -n1)
+        if [ -z "$STAGE1_RESULT_F" ]; then
+            echo "[ablation-grid-layer] missing *_result.json under $STAGE1_OUTPUT_PATH -- has the strength grid finished?" >&2
+            exit 1
+        fi
+        STAGE1_SCORE=$(python -c "
+import json
+with open('$STAGE1_RESULT_F') as f:
+    d = json.load(f)
+print(d['The total score for multiple-choice questions'])
+")
+        if [ -z "$STAGE1_BEST_SCORE" ] || python -c "exit(0 if $STAGE1_SCORE > $STAGE1_BEST_SCORE else 1)"; then
+            STAGE1_BEST_SCORE="$STAGE1_SCORE"
+            S_TEXT="$ST"
+            S_VIS="$SV"
+        fi
+    done
+done
+echo "[ablation-grid-layer] strength winner: s_text=$S_TEXT s_vis=$S_VIS (score=$STAGE1_BEST_SCORE)"
+
+# ── pick the best layer across all 28, strength fixed at (S_TEXT, S_VIS).
+BEST_LAYER=""
+BEST_SCORE=""
+for LAYER in "${LAYERS_ALL[@]}"; do
+    OUTPUT_PATH="results_attnbias_ablation/HuatuoGPT-Vision-7B/sam_both_amplify_noauction_layer${LAYER}_stext${S_TEXT}_svis${S_VIS}"
+    RESULT_F=$(ls "$OUTPUT_PATH"/*_result.json 2>/dev/null | head -n1)
+    if [ -z "$RESULT_F" ]; then
+        echo "[ablation-grid-layer] missing *_result.json under $OUTPUT_PATH -- cannot pick best layer (layer=$LAYER not yet run)" >&2
+        exit 1
+    fi
+    SCORE=$(python -c "
+import json
+with open('$RESULT_F') as f:
+    d = json.load(f)
+print(d['The total score for multiple-choice questions'])
+")
+    echo "[ablation-grid-layer] layer=$LAYER score=$SCORE"
+    if [ -z "$BEST_SCORE" ] || python -c "exit(0 if $SCORE > $BEST_SCORE else 1)"; then
+        BEST_SCORE="$SCORE"
+        BEST_LAYER="$LAYER"
+    fi
+done
+
+echo "[ablation-grid-layer] best overall across all 28 layers: layer=$BEST_LAYER s_text=$S_TEXT s_vis=$S_VIS (score=$BEST_SCORE)"
+echo "[ablation-grid-layer] syncing into every run_eval_attnbias_sam_*.sh via sync_attn_params.py"
+
+python sync_attn_params.py \
+    --layers_text "$BEST_LAYER" --key_bias_text "$S_TEXT" --nonkey_bias_text "0" \
+    --layers_vis  "$BEST_LAYER" --key_bias_vis  "$S_VIS"  --nonkey_bias_vis  "0"
+
+echo "[ablation-grid-layer] done. All run_eval_attnbias_sam_*.sh now use the ablation-selected layer/strength (nonkey_bias=0, amplify_only style)"
+echo "[ablation-grid-layer] (amplify_only keeps its pinned-zero side; text.sh/vis.sh keep their single-modality side -- see sync_attn_params.py)."

HuatuoGPT-Vision-Abla/results_attnbias_ablation/HuatuoGPT-Vision-7B/sam_both_amplify_noauction_layer17_stext3.0_svis2.0/_snapshot/run_ablation_grid_layer_worker.sh

@@ -0,0 +1,144 @@
+#!/bin/bash
+cd "$(dirname "$0")"
+
+# Runs a contiguous slice of the "remaining layers" sweep (the 23 layers NOT
+# already covered by the original LAYERS=(3 10 17 24 27) test) on a single
+# GPU, so the full sweep can be split across GPU 0 and GPU 2 and run
+# concurrently (GPU 1/3 are occupied by other jobs on this workstation) --
+# same pattern as run_ablation_grid_worker.sh for the strength grid, but for
+# layers.
+#
+# NO cache-warmup step here (unlike run_ablation_grid.sh's stage 1): by the
+# time this runs, the Pass0 cache (keyed on data_path/model_path/seed, shared
+# across every combo/script in this repo) has already been fully populated by
+# the earlier strength grid + original 5-layer sweep + run_all_attnbias_sam.sh
+# runs, so both GPU workers should hit pure cache reads with nothing new to
+# write -- eliminating the write-race risk without needing a serialized
+# warmup run first. This assumption only breaks if the cache is somehow still
+# incomplete (e.g. an earlier run was interrupted before finishing Pass 0);
+# if you're not confident the cache is fully warm, warm it manually first
+# (run any single combo alone) before launching this in parallel.
+#
+# S_TEXT/S_VIS are re-derived from the strength grid's results (same as
+# run_ablation_grid_layer.sh), not hand-filled.
+#
+# FINAL METHOD = run_eval_attnbias_sam_both.sh: nonkey_bias_text/vis
+# are pinned at 0 (amplify_only) AND the auction mechanism is OFF, matching
+# run_ablation_grid_worker.sh's strength grid exactly. Output dirs use the
+# "sam_both_amplify_noauction_..." tag to match.
+#
+# Usage:
+#   ./run_ablation_grid_layer_worker.sh <gpu_id> <start_index> <count>
+# Example (23 combos split 12/11 across 2 GPUs):
+#   ./run_ablation_grid_layer_worker.sh 0 0  12   &
+#   ./run_ablation_grid_layer_worker.sh 2 12 11   &
+#   wait
+set -e
+
+GPU_ID="$1"
+START="$2"
+COUNT="$3"
+if [ -z "$GPU_ID" ] || [ -z "$START" ] || [ -z "$COUNT" ]; then
+    echo "usage: $0 <gpu_id> <start_index> <count>" >&2
+    exit 1
+fi
+export CUDA_VISIBLE_DEVICES="$GPU_ID"
+
+source /home/jinhua/miniconda3/etc/profile.d/conda.sh
+conda activate huatuovision
+
+DATA_PATH="/home/jinhua/ExplainableVQA/benchmarks/Medical_Multimodal_Evaluation_Data/medical_multimodel_evaluation_data.json"
+MODEL_PATH="FreedomIntelligence/HuatuoGPT-Vision-7B"
+
+SEED=42
+MAX_NEW_TOKENS=256
+
+MEDSAM3_CONFIG="/home/jinhua/ExplainableVQA/HuatuoGPT-Vision-Abla/MedSAM3/configs/full_lora_config.yaml"
+MEDSAM3_DIR="/home/jinhua/ExplainableVQA/HuatuoGPT-Vision-Abla/MedSAM3"
+
+INTERVENTION="both"
+# the 23 layers NOT in the original LAYERS=(3 10 17 24 27) test -- model has
+# 28 layers total (indices 0-27), confirmed via
+# PretrainedConfig.get_config_dict('FreedomIntelligence/HuatuoGPT-Vision-7B')['num_hidden_layers'] == 28.
+# run_ablation_grid_layer.sh (the driver) does NOT keep its own copy of this
+# array -- it only needs the total count (TOTAL=23) to split indices across
+# GPU 0/1, plus a separate LAYERS_ALL (all 28) for the final "pick best
+# layer" scan. If you change this array's length, update TOTAL/SPLIT in
+# run_ablation_grid_layer.sh to match.
+LAYERS_REMAINING=(0 1 2 4 5 6 7 8 9 11 12 13 14 15 16 18 19 20 21 22 23 25 26)
+STAGE1_LAYER=17
+STRENGTHS_TEXT=(1.0 2.0 3.0 4.0 5.0)
+STRENGTHS_VIS=(1.0 2.0 3.0 4.0 5.0)
+
+MAX_SAMPLES=""
+
+# ── re-derive stage 1's winning (s_text, s_vis) directly from the strength
+#    grid's result files -- same scan-and-compare as run_ablation_grid_pick.sh.
+echo "[ablation-grid-layer-worker gpu=$GPU_ID] re-scanning stage 1's results to find the winning (s_text, s_vis)..."
+S_TEXT=""
+S_VIS=""
+STAGE1_BEST_SCORE=""
+for ST in "${STRENGTHS_TEXT[@]}"; do
+    for SV in "${STRENGTHS_VIS[@]}"; do
+        STAGE1_OUTPUT_PATH="results_attnbias_ablation/HuatuoGPT-Vision-7B/sam_both_amplify_noauction_layer${STAGE1_LAYER}_stext${ST}_svis${SV}"
+        STAGE1_RESULT_F=$(ls "$STAGE1_OUTPUT_PATH"/*_result.json 2>/dev/null | head -n1)
+        if [ -z "$STAGE1_RESULT_F" ]; then
+            echo "[ablation-grid-layer-worker gpu=$GPU_ID] missing *_result.json under $STAGE1_OUTPUT_PATH -- has the strength grid finished?" >&2
+            exit 1
+        fi
+        STAGE1_SCORE=$(python -c "
+import json
+with open('$STAGE1_RESULT_F') as f:
+    d = json.load(f)
+print(d['The total score for multiple-choice questions'])
+")
+        if [ -z "$STAGE1_BEST_SCORE" ] || python -c "exit(0 if $STAGE1_SCORE > $STAGE1_BEST_SCORE else 1)"; then
+            STAGE1_BEST_SCORE="$STAGE1_SCORE"
+            S_TEXT="$ST"
+            S_VIS="$SV"
+        fi
+    done
+done
+echo "[ablation-grid-layer-worker gpu=$GPU_ID] strength winner: s_text=$S_TEXT s_vis=$S_VIS (score=$STAGE1_BEST_SCORE) -- fixing strength at this combo for the layer sweep"
+
+TOTAL=${#LAYERS_REMAINING[@]}
+END=$((START + COUNT))
+if [ "$END" -gt "$TOTAL" ]; then
+    END="$TOTAL"
+fi
+
+echo "[ablation-grid-layer-worker gpu=$GPU_ID] running layers [$START, $END) of $TOTAL total"
+
+for ((i = START; i < END; i++)); do
+    LAYER="${LAYERS_REMAINING[$i]}"
+    OUTPUT_PATH="results_attnbias_ablation/HuatuoGPT-Vision-7B/sam_both_amplify_noauction_layer${LAYER}_stext${S_TEXT}_svis${S_VIS}"
+    if [ -d "$OUTPUT_PATH" ]; then
+        echo "[ablation-grid-layer-worker gpu=$GPU_ID] $OUTPUT_PATH already exists, skipping layer=$LAYER (delete the dir manually to re-run)"
+        continue
+    fi
+
+    echo "=================================================="
+    echo "[ablation-grid-layer-worker gpu=$GPU_ID] idx=$i layer=$LAYER  s_text=$S_TEXT  s_vis=$S_VIS  ->  $OUTPUT_PATH"
+    echo "=================================================="
+
+    accelerate launch --num_processes=1 eval_attnbias_sam.py \
+        --data_path              "$DATA_PATH"              \
+        --model_path             "$MODEL_PATH"             \
+        --seed                   "$SEED"                   \
+        --max_new_tokens         "$MAX_NEW_TOKENS"         \
+        --output_path            "$OUTPUT_PATH"            \
+        --intervention           "$INTERVENTION"           \
+        --attn_layers_text       "$LAYER"                  \
+        --attn_key_bias_text     "$S_TEXT"                 \
+        --attn_nonkey_bias_text  "0"                       \
+        --attn_layers_vis        "$LAYER"                  \
+        --attn_key_bias_vis      "$S_VIS"                  \
+        --attn_nonkey_bias_vis   "0"                       \
+        --medsam3_config         "$MEDSAM3_CONFIG"         \
+        --medsam3_dir            "$MEDSAM3_DIR"            \
+        ${MAX_SAMPLES:+--max_samples "$MAX_SAMPLES"}
+
+    echo "[ablation-grid-layer-worker gpu=$GPU_ID] idx=$i layer=$LAYER finished"
+done
+
+echo "[ablation-grid-layer-worker gpu=$GPU_ID] done with layers [$START, $END)"

HuatuoGPT-Vision-Abla/results_attnbias_ablation/HuatuoGPT-Vision-7B/sam_both_amplify_noauction_layer17_stext3.0_svis2.0/_snapshot/run_ablation_grid_worker.sh

@@ -0,0 +1,116 @@
+#!/bin/bash
+cd "$(dirname "$0")"
+
+# Runs a contiguous slice of the (strength_text, strength_vis) grid on a single
+# GPU, so the full 25-combo sweep can be split across e.g. GPU 0 and GPU 1 and
+# run concurrently. The full pair list is built in the same fixed order
+# (outer S_TEXT, inner S_VIS) as run_ablation_grid_pick.sh expects, so slicing
+# it by [START, START+COUNT) here and picking over the *union* of all slices'
+# output dirs later stays consistent -- this script does NOT do the final
+# "pick best + sync" step, that only makes sense once every slice (on every
+# GPU) has finished; run run_ablation_grid_pick.sh once by hand after all
+# workers exit.
+#
+# FINAL METHOD = run_eval_attnbias_sam_both.sh: nonkey_bias_text/vis
+# are pinned at 0 (amplify_only -- ablation found boosting key regions without
+# suppressing anything else beats full boost+suppress), AND the auction
+# mechanism is OFF (--auction_enable is not passed) -- every Stage1/SAM-detected
+# key position gets boosted, not just an auction-selected subset. This grid
+# searches under the SAME conditions as the final method (no auction), so the
+# winning strength is actually calibrated for what gets used, not for a
+# different (with-auction) config. Output dirs are tagged
+# "sam_both_amplify_noauction_..." so they don't collide with the earlier
+# auction-enabled grid searches' results.
+#
+# Usage:
+#   ./run_ablation_grid_worker.sh <gpu_id> <start_index> <count>
+# Example (25 combos split 13/12 across 2 GPUs -- 0/2 here because GPU 1/3 are
+# occupied by other jobs on this workstation, not the usual 0/1):
+#   CUDA_VISIBLE_DEVICES managed internally from <gpu_id> -- don't set it yourself.
+#   ./run_ablation_grid_worker.sh 0 0  13   &   # combos [0, 13) on GPU 0
+#   ./run_ablation_grid_worker.sh 2 13 12   &   # combos [13, 25) on GPU 2
+#   wait
+#   ./run_ablation_grid_pick.sh
+set -e
+
+GPU_ID="$1"
+START="$2"
+COUNT="$3"
+if [ -z "$GPU_ID" ] || [ -z "$START" ] || [ -z "$COUNT" ]; then
+    echo "usage: $0 <gpu_id> <start_index> <count>" >&2
+    exit 1
+fi
+export CUDA_VISIBLE_DEVICES="$GPU_ID"
+
+source /home/jinhua/miniconda3/etc/profile.d/conda.sh
+conda activate huatuovision
+
+DATA_PATH="/home/jinhua/ExplainableVQA/benchmarks/Medical_Multimodal_Evaluation_Data/medical_multimodel_evaluation_data.json"
+MODEL_PATH="FreedomIntelligence/HuatuoGPT-Vision-7B"
+
+SEED=42
+MAX_NEW_TOKENS=256
+
+MEDSAM3_CONFIG="/home/jinhua/ExplainableVQA/HuatuoGPT-Vision-Abla/MedSAM3/configs/full_lora_config.yaml"
+MEDSAM3_DIR="/home/jinhua/ExplainableVQA/HuatuoGPT-Vision-Abla/MedSAM3"
+
+INTERVENTION="both"
+LAYER=17
+# must stay identical to run_ablation_grid_pick.sh -- these define the pair
+# ordering/indices that <start_index>/<count> slice into.
+STRENGTHS_TEXT=(1.0 2.0 3.0 4.0 5.0)
+STRENGTHS_VIS=(1.0 2.0 3.0 4.0 5.0)
+
+MAX_SAMPLES=""
+
+# flatten into "s_text s_vis" pairs, outer S_TEXT / inner S_VIS -- same order
+# run_ablation_grid_pick.sh iterates in, so index N here means the same combo
+# there.
+PAIRS=()
+for S_TEXT in "${STRENGTHS_TEXT[@]}"; do
+    for S_VIS in "${STRENGTHS_VIS[@]}"; do
+        PAIRS+=("$S_TEXT $S_VIS")
+    done
+done
+
+TOTAL=${#PAIRS[@]}
+END=$((START + COUNT))
+if [ "$END" -gt "$TOTAL" ]; then
+    END="$TOTAL"
+fi
+
+echo "[ablation-grid-worker gpu=$GPU_ID] running combos [$START, $END) of $TOTAL total"
+
+for ((i = START; i < END; i++)); do
+    read -r S_TEXT S_VIS <<< "${PAIRS[$i]}"
+    OUTPUT_PATH="results_attnbias_ablation/HuatuoGPT-Vision-7B/sam_both_amplify_noauction_layer${LAYER}_stext${S_TEXT}_svis${S_VIS}"
+    if [ -d "$OUTPUT_PATH" ]; then
+        echo "[ablation-grid-worker gpu=$GPU_ID] $OUTPUT_PATH already exists, skipping s_text=$S_TEXT s_vis=$S_VIS (delete the dir manually to re-run)"
+        continue
+    fi
+
+    echo "=================================================="
+    echo "[ablation-grid-worker gpu=$GPU_ID] idx=$i layer=$LAYER  s_text=$S_TEXT  s_vis=$S_VIS  ->  $OUTPUT_PATH"
+    echo "=================================================="
+
+    accelerate launch --num_processes=1 eval_attnbias_sam.py \
+        --data_path              "$DATA_PATH"              \
+        --model_path             "$MODEL_PATH"             \
+        --seed                   "$SEED"                   \
+        --max_new_tokens         "$MAX_NEW_TOKENS"         \
+        --output_path            "$OUTPUT_PATH"            \
+        --intervention           "$INTERVENTION"           \
+        --attn_layers_text       "$LAYER"                  \
+        --attn_key_bias_text     "$S_TEXT"                 \
+        --attn_nonkey_bias_text  "0"                       \
+        --attn_layers_vis        "$LAYER"                  \
+        --attn_key_bias_vis      "$S_VIS"                  \
+        --attn_nonkey_bias_vis   "0"                       \
+        --medsam3_config         "$MEDSAM3_CONFIG"         \
+        --medsam3_dir            "$MEDSAM3_DIR"            \
+        ${MAX_SAMPLES:+--max_samples "$MAX_SAMPLES"}
+
+    echo "[ablation-grid-worker gpu=$GPU_ID] idx=$i s_text=$S_TEXT s_vis=$S_VIS finished"
+done
+
+echo "[ablation-grid-worker gpu=$GPU_ID] done with combos [$START, $END)"

HuatuoGPT-Vision-Abla/results_attnbias_ablation/HuatuoGPT-Vision-7B/sam_both_amplify_noauction_layer17_stext3.0_svis2.0/_snapshot/run_all_attnbias_sam.sh

@@ -0,0 +1,36 @@
+#!/bin/bash
+cd "$(dirname "$0")"
+
+# 先跑 baseline（无干预的 eval.py），再依次跑完对比组：
+# none（无attention干预）/ both（正式配置：全量放大关键区域，不开拍卖）/
+# text / vis（单模态消融）/ both_random（真实位置 vs 随机位置消融）
+# run_eval_attnbias_sam_both_noauction.sh 内容跟正式配置both.sh完全一样(both.sh
+# 本身已经改成不开拍卖了)，文件保留但不再放进这个清单，避免重复跑同一件事。
+# 任意一个失败就停止，避免在错误的环境下继续浪费时间
+set -e
+
+# 默认使用 huatuovision conda 环境，不依赖调用者当前激活的环境。
+# source profile.d/conda.sh是必须的：非交互式shell(setsid/nohup这类后台
+# 启动方式)不会自动加载~/.bashrc，conda activate这个shell函数不会被定义，
+# 必须显式source这个脚本才能在当前shell里用conda activate。
+source /home/jinhua/miniconda3/etc/profile.d/conda.sh
+conda activate huatuovision
+
+SCRIPTS=(
+    "run_eval.sh"
+    "run_eval_attnbias_sam_none.sh"
+    "run_eval_attnbias_sam_both.sh"
+    "run_eval_attnbias_sam_text.sh"
+    "run_eval_attnbias_sam_vis.sh"
+    "run_eval_attnbias_sam_both_random.sh"
+)
+
+for script in "${SCRIPTS[@]}"; do
+    echo "=================================================="
+    echo "[run_all] 开始运行 $script"
+    echo "=================================================="
+    bash "$script"
+    echo "[run_all] $script 运行完成"
+done
+
+echo "[run_all] 全部干预模式运行完成"

HuatuoGPT-Vision-Abla/results_attnbias_ablation/HuatuoGPT-Vision-7B/sam_both_amplify_noauction_layer17_stext3.0_svis2.0/_snapshot/run_eval.sh

@@ -0,0 +1,19 @@
+#!/bin/bash
+cd "$(dirname "$0")"
+
+DATA_PATH="/home/jinhua/ExplainableVQA/benchmarks/Medical_Multimodal_Evaluation_Data/medical_multimodel_evaluation_data.json"
+MODEL_PATH="FreedomIntelligence/HuatuoGPT-Vision-7B"
+
+SEED=42
+MAX_NEW_TOKENS=256
+
+# 单卡跑，卡1/卡3被其他任务占用(其他人的进程)，显式锁定卡0(确认空闲的卡)。
+# Single GPU -- pin to GPU 0 explicitly (confirmed free; GPU 1/3 are occupied
+# by other jobs on this workstation).
+export CUDA_VISIBLE_DEVICES=0
+
+accelerate launch --num_processes=1 eval.py \
+    --data_path       "$DATA_PATH"       \
+    --model_path      "$MODEL_PATH"      \
+    --seed            "$SEED"            \
+    --max_new_tokens  "$MAX_NEW_TOKENS"

HuatuoGPT-Vision-Abla/results_attnbias_ablation/HuatuoGPT-Vision-7B/sam_both_amplify_noauction_layer17_stext3.0_svis2.0/_snapshot/run_eval_attnbias_sam_both.sh

@@ -0,0 +1,45 @@
+#!/bin/bash
+cd "$(dirname "$0")"
+
+DATA_PATH="/home/jinhua/ExplainableVQA/benchmarks/Medical_Multimodal_Evaluation_Data/medical_multimodel_evaluation_data.json"
+MODEL_PATH="FreedomIntelligence/HuatuoGPT-Vision-7B"
+OUTPUT_PATH="results_attnbias/HuatuoGPT-Vision-7B/sam_both"
+
+SEED=42
+MAX_NEW_TOKENS=256
+
+MEDSAM3_CONFIG="/home/jinhua/ExplainableVQA/HuatuoGPT-Vision-Abla/MedSAM3/configs/full_lora_config.yaml"
+MEDSAM3_DIR="/home/jinhua/ExplainableVQA/HuatuoGPT-Vision-Abla/MedSAM3"
+
+# FINAL METHOD: both mode, amplify-only (nonkey_bias pinned at 0 -- boost key
+# regions, don't suppress anything else) and auction OFF -- every Stage1/SAM-
+# detected key position gets boosted directly, not just an auction-selected
+# subset (ablation found the auction mechanism doesn't help over this simpler
+# design; see run_eval_attnbias_sam_both_noauction.sh, kept around for
+# reference but no longer part of the active comparison set).
+INTERVENTION="both"
+ATTN_LAYERS_TEXT=17
+ATTN_KEY_BIAS_TEXT=5.0
+ATTN_NONKEY_BIAS_TEXT=0.0
+ATTN_LAYERS_VIS=17
+ATTN_KEY_BIAS_VIS=1.0
+ATTN_NONKEY_BIAS_VIS=0.0
+
+MAX_SAMPLES=""
+
+accelerate launch --num_processes=1 eval_attnbias_sam.py \
+    --data_path                "$DATA_PATH"                \
+    --model_path               "$MODEL_PATH"               \
+    --seed                     "$SEED"                     \
+    --max_new_tokens           "$MAX_NEW_TOKENS"           \
+    --output_path              "$OUTPUT_PATH"              \
+    --intervention              "$INTERVENTION"             \
+    --attn_layers_text          "$ATTN_LAYERS_TEXT"         \
+    --attn_key_bias_text        "$ATTN_KEY_BIAS_TEXT"       \
+    --attn_nonkey_bias_text     "$ATTN_NONKEY_BIAS_TEXT"    \
+    --attn_layers_vis           "$ATTN_LAYERS_VIS"          \
+    --attn_key_bias_vis         "$ATTN_KEY_BIAS_VIS"        \
+    --attn_nonkey_bias_vis      "$ATTN_NONKEY_BIAS_VIS"     \
+    --medsam3_config            "$MEDSAM3_CONFIG"           \
+    --medsam3_dir               "$MEDSAM3_DIR"              \
+    ${MAX_SAMPLES:+--max_samples "$MAX_SAMPLES"}

HuatuoGPT-Vision-Abla/results_attnbias_ablation/HuatuoGPT-Vision-7B/sam_both_amplify_noauction_layer17_stext3.0_svis2.0/_snapshot/run_eval_attnbias_sam_both_amplify_only.sh

@@ -0,0 +1,52 @@
+#!/bin/bash
+cd "$(dirname "$0")"
+
+DATA_PATH="/home/jinhua/ExplainableVQA/benchmarks/Medical_Multimodal_Evaluation_Data/medical_multimodel_evaluation_data.json"
+MODEL_PATH="FreedomIntelligence/HuatuoGPT-Vision-7B"
+OUTPUT_PATH="results_attnbias/HuatuoGPT-Vision-7B/sam_both_amplify_only"
+
+SEED=42
+MAX_NEW_TOKENS=256
+
+MEDSAM3_CONFIG="/home/jinhua/ExplainableVQA/HuatuoGPT-Vision-Abla/MedSAM3/configs/full_lora_config.yaml"
+MEDSAM3_DIR="/home/jinhua/ExplainableVQA/HuatuoGPT-Vision-Abla/MedSAM3"
+
+# Ablation: amplify-only. Key positions still get the positive bias, but
+# nonkey_bias_* is 0 instead of negative -- isolates how much of the "both"
+# result comes from amplifying keyword text / SAM patches, with no suppression
+# of distractors. Auction stays ON with the same ratio/alpha/beta as
+# run_eval_attnbias_sam_both.sh, so this differs from "both" by exactly one
+# factor (nonkey_bias_*); compare against run_eval_attnbias_sam_both_suppress_only.sh
+# (same setup, opposite side zeroed) and run_eval_attnbias_sam_both.sh.
+INTERVENTION="both"
+ATTN_LAYERS_TEXT=17
+ATTN_KEY_BIAS_TEXT=5.0
+ATTN_NONKEY_BIAS_TEXT=0.0
+ATTN_LAYERS_VIS=17
+ATTN_KEY_BIAS_VIS=1.0
+ATTN_NONKEY_BIAS_VIS=0.0
+
+AUCTION_TOTAL_RATIO=0.8
+AUCTION_NEG_RATIO=0.8
+
+MAX_SAMPLES=""
+
+accelerate launch --num_processes=1 eval_attnbias_sam.py \
+    --data_path                "$DATA_PATH"                \
+    --model_path               "$MODEL_PATH"               \
+    --seed                     "$SEED"                     \
+    --max_new_tokens           "$MAX_NEW_TOKENS"           \
+    --output_path              "$OUTPUT_PATH"              \
+    --intervention              "$INTERVENTION"             \
+    --attn_layers_text          "$ATTN_LAYERS_TEXT"         \
+    --attn_key_bias_text        "$ATTN_KEY_BIAS_TEXT"       \
+    --attn_nonkey_bias_text     "$ATTN_NONKEY_BIAS_TEXT"    \
+    --attn_layers_vis           "$ATTN_LAYERS_VIS"          \
+    --attn_key_bias_vis         "$ATTN_KEY_BIAS_VIS"        \
+    --attn_nonkey_bias_vis      "$ATTN_NONKEY_BIAS_VIS"     \
+    --medsam3_config            "$MEDSAM3_CONFIG"           \
+    --medsam3_dir               "$MEDSAM3_DIR"              \
+    --auction_enable                                       \
+    --auction_total_ratio       "$AUCTION_TOTAL_RATIO"      \
+    --auction_neg_ratio         "$AUCTION_NEG_RATIO"        \
+    ${MAX_SAMPLES:+--max_samples "$MAX_SAMPLES"}

HuatuoGPT-Vision-Abla/results_attnbias_ablation/HuatuoGPT-Vision-7B/sam_both_amplify_noauction_layer17_stext3.0_svis2.0/_snapshot/run_eval_attnbias_sam_both_noauction.sh

@@ -0,0 +1,42 @@
+#!/bin/bash
+cd "$(dirname "$0")"
+
+DATA_PATH="/home/jinhua/ExplainableVQA/benchmarks/Medical_Multimodal_Evaluation_Data/medical_multimodel_evaluation_data.json"
+MODEL_PATH="FreedomIntelligence/HuatuoGPT-Vision-7B"
+OUTPUT_PATH="results_attnbias/HuatuoGPT-Vision-7B/sam_both_noauction"
+
+SEED=42
+MAX_NEW_TOKENS=256
+
+MEDSAM3_CONFIG="/home/jinhua/ExplainableVQA/HuatuoGPT-Vision-Abla/MedSAM3/configs/full_lora_config.yaml"
+MEDSAM3_DIR="/home/jinhua/ExplainableVQA/HuatuoGPT-Vision-Abla/MedSAM3"
+
+# both mode, auction OFF (legacy behavior): every SAM patch / scispaCy keyword
+# gets amplified, everything else gets suppressed -- no scoring, no budget. This
+# is the "before" half of the main ablation: none vs both-noauction vs both+auction.
+INTERVENTION="both"
+ATTN_LAYERS_TEXT=17
+ATTN_KEY_BIAS_TEXT=5.0
+ATTN_NONKEY_BIAS_TEXT=0.0
+ATTN_LAYERS_VIS=17
+ATTN_KEY_BIAS_VIS=1.0
+ATTN_NONKEY_BIAS_VIS=0.0
+
+MAX_SAMPLES=""
+
+accelerate launch --num_processes=1 eval_attnbias_sam.py \
+    --data_path              "$DATA_PATH"              \
+    --model_path             "$MODEL_PATH"             \
+    --seed                   "$SEED"                   \
+    --max_new_tokens         "$MAX_NEW_TOKENS"         \
+    --output_path            "$OUTPUT_PATH"            \
+    --intervention           "$INTERVENTION"           \
+    --attn_layers_text       "$ATTN_LAYERS_TEXT"       \
+    --attn_key_bias_text     "$ATTN_KEY_BIAS_TEXT"     \
+    --attn_nonkey_bias_text  "$ATTN_NONKEY_BIAS_TEXT"  \
+    --attn_layers_vis        "$ATTN_LAYERS_VIS"        \
+    --attn_key_bias_vis      "$ATTN_KEY_BIAS_VIS"      \
+    --attn_nonkey_bias_vis   "$ATTN_NONKEY_BIAS_VIS"   \
+    --medsam3_config         "$MEDSAM3_CONFIG"         \
+    --medsam3_dir            "$MEDSAM3_DIR"            \
+    ${MAX_SAMPLES:+--max_samples "$MAX_SAMPLES"}

HuatuoGPT-Vision-Abla/results_attnbias_ablation/HuatuoGPT-Vision-7B/sam_both_amplify_noauction_layer17_stext3.0_svis2.0/_snapshot/run_eval_attnbias_sam_both_random.sh

@@ -0,0 +1,51 @@
+#!/bin/bash
+cd "$(dirname "$0")"
+
+DATA_PATH="/home/jinhua/ExplainableVQA/benchmarks/Medical_Multimodal_Evaluation_Data/medical_multimodel_evaluation_data.json"
+MODEL_PATH="FreedomIntelligence/HuatuoGPT-Vision-7B"
+OUTPUT_PATH="results_attnbias/HuatuoGPT-Vision-7B/sam_both_random"
+
+SEED=42
+MAX_NEW_TOKENS=256
+
+MEDSAM3_CONFIG="/home/jinhua/ExplainableVQA/HuatuoGPT-Vision-Abla/MedSAM3/configs/full_lora_config.yaml"
+MEDSAM3_DIR="/home/jinhua/ExplainableVQA/HuatuoGPT-Vision-Abla/MedSAM3"
+
+# Ablation control: same "both" bias strengths/layers as
+# run_eval_attnbias_sam_both.sh (amplify-only, auction OFF), but the scispaCy
+# keyword positions and SAM bbox patches are each replaced
+# (--ablation_random_text/--ablation_random_vis) with an equal-count RANDOM
+# sample from the same candidate pool (all question text tokens / full 24x24
+# visual grid, excluding the true key positions), seeded by $SEED for
+# reproducibility. This differs from the final method by exactly one factor
+# (random vs real key positions) -- tests whether gains come from biasing the
+# *semantically correct* tokens, or merely from biasing *some* tokens of the
+# same count. Compare this run's accuracy against run_eval_attnbias_sam_both.sh.
+INTERVENTION="both"
+ATTN_LAYERS_TEXT=17
+ATTN_KEY_BIAS_TEXT=5.0
+ATTN_NONKEY_BIAS_TEXT=0.0
+ATTN_LAYERS_VIS=17
+ATTN_KEY_BIAS_VIS=1.0
+ATTN_NONKEY_BIAS_VIS=0.0
+
+MAX_SAMPLES=""
+
+accelerate launch --num_processes=1 eval_attnbias_sam.py \
+    --data_path                "$DATA_PATH"                \
+    --model_path               "$MODEL_PATH"               \
+    --seed                     "$SEED"                     \
+    --max_new_tokens           "$MAX_NEW_TOKENS"           \
+    --output_path              "$OUTPUT_PATH"              \
+    --intervention              "$INTERVENTION"             \
+    --attn_layers_text          "$ATTN_LAYERS_TEXT"         \
+    --attn_key_bias_text        "$ATTN_KEY_BIAS_TEXT"       \
+    --attn_nonkey_bias_text     "$ATTN_NONKEY_BIAS_TEXT"    \
+    --attn_layers_vis           "$ATTN_LAYERS_VIS"          \
+    --attn_key_bias_vis         "$ATTN_KEY_BIAS_VIS"        \
+    --attn_nonkey_bias_vis      "$ATTN_NONKEY_BIAS_VIS"     \
+    --medsam3_config            "$MEDSAM3_CONFIG"           \
+    --medsam3_dir               "$MEDSAM3_DIR"              \
+    --ablation_random_text                                 \
+    --ablation_random_vis                                  \
+    ${MAX_SAMPLES:+--max_samples "$MAX_SAMPLES"}

HuatuoGPT-Vision-Abla/results_attnbias_ablation/HuatuoGPT-Vision-7B/sam_both_amplify_noauction_layer17_stext3.0_svis2.0/_snapshot/run_eval_attnbias_sam_both_suppress_only.sh

@@ -0,0 +1,53 @@
+#!/bin/bash
+cd "$(dirname "$0")"
+
+DATA_PATH="/home/jinhua/ExplainableVQA/benchmarks/Medical_Multimodal_Evaluation_Data/medical_multimodel_evaluation_data.json"
+MODEL_PATH="FreedomIntelligence/HuatuoGPT-Vision-7B"
+OUTPUT_PATH="results_attnbias/HuatuoGPT-Vision-7B/sam_both_suppress_only"
+
+SEED=42
+MAX_NEW_TOKENS=256
+
+MEDSAM3_CONFIG="/home/jinhua/ExplainableVQA/HuatuoGPT-Vision-Abla/MedSAM3/configs/full_lora_config.yaml"
+MEDSAM3_DIR="/home/jinhua/ExplainableVQA/HuatuoGPT-Vision-Abla/MedSAM3"
+
+# Ablation: suppress-only. nonkey_bias_* keeps the negative bias (distractors
+# suppressed), but key_bias_* is 0 instead of positive -- isolates how much of
+# the "both" result comes from suppressing non-keyword text / non-SAM visual
+# patches, with no amplification of the key set. Auction stays ON with the
+# same ratio/alpha/beta as run_eval_attnbias_sam_both.sh, so this differs from
+# "both" by exactly one factor (key_bias_*); compare against
+# run_eval_attnbias_sam_both_amplify_only.sh (same setup, opposite side
+# zeroed) and run_eval_attnbias_sam_both.sh.
+INTERVENTION="both"
+ATTN_LAYERS_TEXT=17
+ATTN_KEY_BIAS_TEXT=0.0
+ATTN_NONKEY_BIAS_TEXT=-3.0
+ATTN_LAYERS_VIS=17
+ATTN_KEY_BIAS_VIS=0.0
+ATTN_NONKEY_BIAS_VIS=-3.0
+
+AUCTION_TOTAL_RATIO=0.8
+AUCTION_NEG_RATIO=0.8
+
+MAX_SAMPLES=""
+
+accelerate launch --num_processes=1 eval_attnbias_sam.py \
+    --data_path                "$DATA_PATH"                \
+    --model_path               "$MODEL_PATH"               \
+    --seed                     "$SEED"                     \
+    --max_new_tokens           "$MAX_NEW_TOKENS"           \
+    --output_path              "$OUTPUT_PATH"              \
+    --intervention              "$INTERVENTION"             \
+    --attn_layers_text          "$ATTN_LAYERS_TEXT"         \
+    --attn_key_bias_text        "$ATTN_KEY_BIAS_TEXT"       \
+    --attn_nonkey_bias_text     "$ATTN_NONKEY_BIAS_TEXT"    \
+    --attn_layers_vis           "$ATTN_LAYERS_VIS"          \
+    --attn_key_bias_vis         "$ATTN_KEY_BIAS_VIS"        \
+    --attn_nonkey_bias_vis      "$ATTN_NONKEY_BIAS_VIS"     \
+    --medsam3_config            "$MEDSAM3_CONFIG"           \
+    --medsam3_dir               "$MEDSAM3_DIR"              \
+    --auction_enable                                       \
+    --auction_total_ratio       "$AUCTION_TOTAL_RATIO"      \
+    --auction_neg_ratio         "$AUCTION_NEG_RATIO"        \
+    ${MAX_SAMPLES:+--max_samples "$MAX_SAMPLES"}

HuatuoGPT-Vision-Abla/results_attnbias_ablation/HuatuoGPT-Vision-7B/sam_both_amplify_noauction_layer17_stext3.0_svis2.0/_snapshot/run_eval_attnbias_sam_none.sh

@@ -0,0 +1,40 @@
+#!/bin/bash
+cd "$(dirname "$0")"
+
+DATA_PATH="/home/jinhua/ExplainableVQA/benchmarks/Medical_Multimodal_Evaluation_Data/medical_multimodel_evaluation_data.json"
+MODEL_PATH="FreedomIntelligence/HuatuoGPT-Vision-7B"
+OUTPUT_PATH="results_attnbias/HuatuoGPT-Vision-7B/sam_none"
+
+SEED=42
+MAX_NEW_TOKENS=256
+
+MEDSAM3_CONFIG="/home/jinhua/ExplainableVQA/HuatuoGPT-Vision-Abla/MedSAM3/configs/full_lora_config.yaml"
+MEDSAM3_DIR="/home/jinhua/ExplainableVQA/HuatuoGPT-Vision-Abla/MedSAM3"
+
+# none: no attention intervention (model_output reuses plain inference result)
+INTERVENTION="none"
+ATTN_LAYERS_TEXT="17"
+ATTN_KEY_BIAS_TEXT=3.0
+ATTN_NONKEY_BIAS_TEXT=-3.0
+ATTN_LAYERS_VIS="17"
+ATTN_KEY_BIAS_VIS=3.0
+ATTN_NONKEY_BIAS_VIS=-3.0
+
+MAX_SAMPLES=""
+
+accelerate launch --num_processes=1 eval_attnbias_sam.py \
+    --data_path              "$DATA_PATH"              \
+    --model_path             "$MODEL_PATH"             \
+    --seed                   "$SEED"                   \
+    --max_new_tokens         "$MAX_NEW_TOKENS"         \
+    --output_path            "$OUTPUT_PATH"            \
+    --intervention           "$INTERVENTION"           \
+    --attn_layers_text       "$ATTN_LAYERS_TEXT"       \
+    --attn_key_bias_text     "$ATTN_KEY_BIAS_TEXT"     \
+    --attn_nonkey_bias_text  "$ATTN_NONKEY_BIAS_TEXT"  \
+    --attn_layers_vis        "$ATTN_LAYERS_VIS"        \
+    --attn_key_bias_vis      "$ATTN_KEY_BIAS_VIS"      \
+    --attn_nonkey_bias_vis   "$ATTN_NONKEY_BIAS_VIS"   \
+    --medsam3_config         "$MEDSAM3_CONFIG"         \
+    --medsam3_dir            "$MEDSAM3_DIR"            \
+    ${MAX_SAMPLES:+--max_samples "$MAX_SAMPLES"}

HuatuoGPT-Vision-Abla/results_attnbias_ablation/HuatuoGPT-Vision-7B/sam_both_amplify_noauction_layer17_stext3.0_svis2.0/_snapshot/run_eval_attnbias_sam_text.sh

@@ -0,0 +1,38 @@
+#!/bin/bash
+cd "$(dirname "$0")"
+
+DATA_PATH="/home/jinhua/ExplainableVQA/benchmarks/Medical_Multimodal_Evaluation_Data/medical_multimodel_evaluation_data.json"
+MODEL_PATH="FreedomIntelligence/HuatuoGPT-Vision-7B"
+OUTPUT_PATH="results_attnbias/HuatuoGPT-Vision-7B/sam_text"
+
+SEED=42
+MAX_NEW_TOKENS=256
+
+# Modality ablation: text-only. Only the scispaCy keyword text positions get
+# biased (key_bias_text / nonkey_bias_text); the vis side is forced to 0 bias
+# inside _make_attn_ctx regardless of what's passed here. SAM/MedSAM3 is
+# skipped entirely (_needs_sam is False for "text"), so no medsam3_* args.
+# Matches run_eval_attnbias_sam_both.sh's mechanism (amplify-only, auction
+# OFF) so this differs from the final method by exactly one factor (modality
+# restricted to text); compare against run_eval_attnbias_sam_vis.sh and
+# run_eval_attnbias_sam_both.sh.
+INTERVENTION="text"
+ATTN_LAYERS_TEXT=17
+ATTN_KEY_BIAS_TEXT=5.0
+ATTN_NONKEY_BIAS_TEXT=0.0
+ATTN_LAYERS_VIS="17"
+
+MAX_SAMPLES=""
+
+accelerate launch --num_processes=1 eval_attnbias_sam.py \
+    --data_path                "$DATA_PATH"                \
+    --model_path               "$MODEL_PATH"               \
+    --seed                     "$SEED"                     \
+    --max_new_tokens           "$MAX_NEW_TOKENS"           \
+    --output_path              "$OUTPUT_PATH"              \
+    --intervention              "$INTERVENTION"             \
+    --attn_layers_text          "$ATTN_LAYERS_TEXT"         \
+    --attn_key_bias_text        "$ATTN_KEY_BIAS_TEXT"       \
+    --attn_nonkey_bias_text     "$ATTN_NONKEY_BIAS_TEXT"    \
+    --attn_layers_vis           "$ATTN_LAYERS_VIS"          \
+    ${MAX_SAMPLES:+--max_samples "$MAX_SAMPLES"}

HuatuoGPT-Vision-Abla/results_attnbias_ablation/HuatuoGPT-Vision-7B/sam_both_amplify_noauction_layer17_stext3.0_svis2.0/_snapshot/run_eval_attnbias_sam_vis.sh

@@ -0,0 +1,42 @@
+#!/bin/bash
+cd "$(dirname "$0")"
+
+DATA_PATH="/home/jinhua/ExplainableVQA/benchmarks/Medical_Multimodal_Evaluation_Data/medical_multimodel_evaluation_data.json"
+MODEL_PATH="FreedomIntelligence/HuatuoGPT-Vision-7B"
+OUTPUT_PATH="results_attnbias/HuatuoGPT-Vision-7B/sam_vis"
+
+SEED=42
+MAX_NEW_TOKENS=256
+
+MEDSAM3_CONFIG="/home/jinhua/ExplainableVQA/HuatuoGPT-Vision-Abla/MedSAM3/configs/full_lora_config.yaml"
+MEDSAM3_DIR="/home/jinhua/ExplainableVQA/HuatuoGPT-Vision-Abla/MedSAM3"
+
+# Modality ablation: vis-only. Only the SAM bbox patches get biased
+# (key_bias_vis / nonkey_bias_vis); the text side is forced to 0 bias inside
+# _make_attn_ctx regardless of what's passed here. SAM/MedSAM3 still runs
+# (_needs_sam is True for "vis"). Matches run_eval_attnbias_sam_both.sh's
+# mechanism (amplify-only, auction OFF) so this differs from the final method
+# by exactly one factor (modality restricted to vis); compare against
+# run_eval_attnbias_sam_text.sh and run_eval_attnbias_sam_both.sh.
+INTERVENTION="vis"
+ATTN_LAYERS_TEXT=17
+ATTN_LAYERS_VIS=17
+ATTN_KEY_BIAS_VIS=1.0
+ATTN_NONKEY_BIAS_VIS=0.0
+
+MAX_SAMPLES=""
+
+accelerate launch --num_processes=1 eval_attnbias_sam.py \
+    --data_path                "$DATA_PATH"                \
+    --model_path               "$MODEL_PATH"               \
+    --seed                     "$SEED"                     \
+    --max_new_tokens           "$MAX_NEW_TOKENS"           \
+    --output_path              "$OUTPUT_PATH"              \
+    --intervention              "$INTERVENTION"             \
+    --attn_layers_text          "$ATTN_LAYERS_TEXT"         \
+    --attn_layers_vis           "$ATTN_LAYERS_VIS"          \
+    --attn_key_bias_vis         "$ATTN_KEY_BIAS_VIS"        \
+    --attn_nonkey_bias_vis      "$ATTN_NONKEY_BIAS_VIS"     \
+    --medsam3_config            "$MEDSAM3_CONFIG"           \
+    --medsam3_dir               "$MEDSAM3_DIR"              \
+    ${MAX_SAMPLES:+--max_samples "$MAX_SAMPLES"}

HuatuoGPT-Vision-Abla/results_attnbias_ablation/HuatuoGPT-Vision-7B/sam_both_amplify_noauction_layer17_stext3.0_svis3.0/HuatuoGPT-Vision-7B_medical_multimodel_evaluation_data_both_Lt17_kt3.0_nt0.0_Lv17_kv3.0_nv0.0_result.json

@@ -0,0 +1,22 @@
+{
+  "OmniMedVQA": 0.7461344839985616,
+  "PMC-VQA_test": 0.548,
+  "PathVQA_test": 0.5809042236763831,
+  "SLAKE_test": 0.7692307692307693,
+  "VQA-RAD_test": 0.6334661354581673,
+  "MMMU_Medical_Validation": 0.5310344827586206,
+  "The total score for multiple-choice questions": 0.6882298531622153,
+  "OmniMedVQA_MR (Mag-netic Resonance Imaging)": 0.7287128712871287,
+  "OmniMedVQA_ultrasound": 0.7776957163958641,
+  "OmniMedVQA_OCT (Optical Coherence Tomography": 0.8666666666666667,
+  "OmniMedVQA_Dermoscopy": 0.7211055276381909,
+  "OmniMedVQA_CT(Computed Tomography)": 0.6685082872928176,
+  "OmniMedVQA_X-Ray": 0.8056918547595682,
+  "OmniMedVQA_Microscopy Images": 0.7870879120879121,
+  "OmniMedVQA_Fundus Photography": 0.8146964856230032,
+  "MMMU_Medical_Validation_validation_Public_Health": 0.5,
+  "MMMU_Medical_Validation_validation_Diagnostics_and_Laboratory_Medicine": 0.4,
+  "MMMU_Medical_Validation_validation_Pharmacy": 0.48148148148148145,
+  "MMMU_Medical_Validation_validation_Basic_Medical_Science": 0.6071428571428571,
+  "MMMU_Medical_Validation_validation_Clinical_Medicine": 0.6666666666666666
+}

HuatuoGPT-Vision-Abla/results_attnbias_ablation/HuatuoGPT-Vision-7B/sam_both_amplify_noauction_layer17_stext3.0_svis3.0/_snapshot/run_ablation_grid_layer_worker.sh

@@ -0,0 +1,144 @@
+#!/bin/bash
+cd "$(dirname "$0")"
+
+# Runs a contiguous slice of the "remaining layers" sweep (the 23 layers NOT
+# already covered by the original LAYERS=(3 10 17 24 27) test) on a single
+# GPU, so the full sweep can be split across GPU 0 and GPU 2 and run
+# concurrently (GPU 1/3 are occupied by other jobs on this workstation) --
+# same pattern as run_ablation_grid_worker.sh for the strength grid, but for
+# layers.
+#
+# NO cache-warmup step here (unlike run_ablation_grid.sh's stage 1): by the
+# time this runs, the Pass0 cache (keyed on data_path/model_path/seed, shared
+# across every combo/script in this repo) has already been fully populated by
+# the earlier strength grid + original 5-layer sweep + run_all_attnbias_sam.sh
+# runs, so both GPU workers should hit pure cache reads with nothing new to
+# write -- eliminating the write-race risk without needing a serialized
+# warmup run first. This assumption only breaks if the cache is somehow still
+# incomplete (e.g. an earlier run was interrupted before finishing Pass 0);
+# if you're not confident the cache is fully warm, warm it manually first
+# (run any single combo alone) before launching this in parallel.
+#
+# S_TEXT/S_VIS are re-derived from the strength grid's results (same as
+# run_ablation_grid_layer.sh), not hand-filled.
+#
+# FINAL METHOD = run_eval_attnbias_sam_both.sh: nonkey_bias_text/vis
+# are pinned at 0 (amplify_only) AND the auction mechanism is OFF, matching
+# run_ablation_grid_worker.sh's strength grid exactly. Output dirs use the
+# "sam_both_amplify_noauction_..." tag to match.
+#
+# Usage:
+#   ./run_ablation_grid_layer_worker.sh <gpu_id> <start_index> <count>
+# Example (23 combos split 12/11 across 2 GPUs):
+#   ./run_ablation_grid_layer_worker.sh 0 0  12   &
+#   ./run_ablation_grid_layer_worker.sh 2 12 11   &
+#   wait
+set -e
+
+GPU_ID="$1"
+START="$2"
+COUNT="$3"
+if [ -z "$GPU_ID" ] || [ -z "$START" ] || [ -z "$COUNT" ]; then
+    echo "usage: $0 <gpu_id> <start_index> <count>" >&2
+    exit 1
+fi
+export CUDA_VISIBLE_DEVICES="$GPU_ID"
+
+source /home/jinhua/miniconda3/etc/profile.d/conda.sh
+conda activate huatuovision
+
+DATA_PATH="/home/jinhua/ExplainableVQA/benchmarks/Medical_Multimodal_Evaluation_Data/medical_multimodel_evaluation_data.json"
+MODEL_PATH="FreedomIntelligence/HuatuoGPT-Vision-7B"
+
+SEED=42
+MAX_NEW_TOKENS=256
+
+MEDSAM3_CONFIG="/home/jinhua/ExplainableVQA/HuatuoGPT-Vision-Abla/MedSAM3/configs/full_lora_config.yaml"
+MEDSAM3_DIR="/home/jinhua/ExplainableVQA/HuatuoGPT-Vision-Abla/MedSAM3"
+
+INTERVENTION="both"
+# the 23 layers NOT in the original LAYERS=(3 10 17 24 27) test -- model has
+# 28 layers total (indices 0-27), confirmed via
+# PretrainedConfig.get_config_dict('FreedomIntelligence/HuatuoGPT-Vision-7B')['num_hidden_layers'] == 28.
+# run_ablation_grid_layer.sh (the driver) does NOT keep its own copy of this
+# array -- it only needs the total count (TOTAL=23) to split indices across
+# GPU 0/1, plus a separate LAYERS_ALL (all 28) for the final "pick best
+# layer" scan. If you change this array's length, update TOTAL/SPLIT in
+# run_ablation_grid_layer.sh to match.
+LAYERS_REMAINING=(0 1 2 4 5 6 7 8 9 11 12 13 14 15 16 18 19 20 21 22 23 25 26)
+STAGE1_LAYER=17
+STRENGTHS_TEXT=(1.0 2.0 3.0 4.0 5.0)
+STRENGTHS_VIS=(1.0 2.0 3.0 4.0 5.0)
+
+MAX_SAMPLES=""
+
+# ── re-derive stage 1's winning (s_text, s_vis) directly from the strength
+#    grid's result files -- same scan-and-compare as run_ablation_grid_pick.sh.
+echo "[ablation-grid-layer-worker gpu=$GPU_ID] re-scanning stage 1's results to find the winning (s_text, s_vis)..."
+S_TEXT=""
+S_VIS=""
+STAGE1_BEST_SCORE=""
+for ST in "${STRENGTHS_TEXT[@]}"; do
+    for SV in "${STRENGTHS_VIS[@]}"; do
+        STAGE1_OUTPUT_PATH="results_attnbias_ablation/HuatuoGPT-Vision-7B/sam_both_amplify_noauction_layer${STAGE1_LAYER}_stext${ST}_svis${SV}"
+        STAGE1_RESULT_F=$(ls "$STAGE1_OUTPUT_PATH"/*_result.json 2>/dev/null | head -n1)
+        if [ -z "$STAGE1_RESULT_F" ]; then
+            echo "[ablation-grid-layer-worker gpu=$GPU_ID] missing *_result.json under $STAGE1_OUTPUT_PATH -- has the strength grid finished?" >&2
+            exit 1
+        fi
+        STAGE1_SCORE=$(python -c "
+import json
+with open('$STAGE1_RESULT_F') as f:
+    d = json.load(f)
+print(d['The total score for multiple-choice questions'])
+")
+        if [ -z "$STAGE1_BEST_SCORE" ] || python -c "exit(0 if $STAGE1_SCORE > $STAGE1_BEST_SCORE else 1)"; then
+            STAGE1_BEST_SCORE="$STAGE1_SCORE"
+            S_TEXT="$ST"
+            S_VIS="$SV"
+        fi
+    done
+done
+echo "[ablation-grid-layer-worker gpu=$GPU_ID] strength winner: s_text=$S_TEXT s_vis=$S_VIS (score=$STAGE1_BEST_SCORE) -- fixing strength at this combo for the layer sweep"
+
+TOTAL=${#LAYERS_REMAINING[@]}
+END=$((START + COUNT))
+if [ "$END" -gt "$TOTAL" ]; then
+    END="$TOTAL"
+fi
+
+echo "[ablation-grid-layer-worker gpu=$GPU_ID] running layers [$START, $END) of $TOTAL total"
+
+for ((i = START; i < END; i++)); do
+    LAYER="${LAYERS_REMAINING[$i]}"
+    OUTPUT_PATH="results_attnbias_ablation/HuatuoGPT-Vision-7B/sam_both_amplify_noauction_layer${LAYER}_stext${S_TEXT}_svis${S_VIS}"
+    if [ -d "$OUTPUT_PATH" ]; then
+        echo "[ablation-grid-layer-worker gpu=$GPU_ID] $OUTPUT_PATH already exists, skipping layer=$LAYER (delete the dir manually to re-run)"
+        continue
+    fi
+
+    echo "=================================================="
+    echo "[ablation-grid-layer-worker gpu=$GPU_ID] idx=$i layer=$LAYER  s_text=$S_TEXT  s_vis=$S_VIS  ->  $OUTPUT_PATH"
+    echo "=================================================="
+
+    accelerate launch --num_processes=1 eval_attnbias_sam.py \
+        --data_path              "$DATA_PATH"              \
+        --model_path             "$MODEL_PATH"             \
+        --seed                   "$SEED"                   \
+        --max_new_tokens         "$MAX_NEW_TOKENS"         \
+        --output_path            "$OUTPUT_PATH"            \
+        --intervention           "$INTERVENTION"           \
+        --attn_layers_text       "$LAYER"                  \
+        --attn_key_bias_text     "$S_TEXT"                 \
+        --attn_nonkey_bias_text  "0"                       \
+        --attn_layers_vis        "$LAYER"                  \
+        --attn_key_bias_vis      "$S_VIS"                  \
+        --attn_nonkey_bias_vis   "0"                       \
+        --medsam3_config         "$MEDSAM3_CONFIG"         \
+        --medsam3_dir            "$MEDSAM3_DIR"            \
+        ${MAX_SAMPLES:+--max_samples "$MAX_SAMPLES"}
+
+    echo "[ablation-grid-layer-worker gpu=$GPU_ID] idx=$i layer=$LAYER finished"
+done
+
+echo "[ablation-grid-layer-worker gpu=$GPU_ID] done with layers [$START, $END)"

HuatuoGPT-Vision-Abla/results_attnbias_ablation/HuatuoGPT-Vision-7B/sam_both_amplify_noauction_layer17_stext3.0_svis3.0/_snapshot/run_ablation_grid_pick.sh

@@ -0,0 +1,68 @@
+#!/bin/bash
+cd "$(dirname "$0")"
+
+# Picks the best (strength_text, strength_vis) across the full grid and syncs
+# it into every run_eval_attnbias_sam_*.sh via sync_attn_params.py. Run this
+# ONCE, by hand, after every run_ablation_grid_worker.sh slice (on every GPU
+# you split the grid across) has finished -- it requires *all* combos'
+# *_result.json to be present, since the "best" choice isn't meaningful until
+# the whole grid has been scored.
+#
+# FINAL METHOD = run_eval_attnbias_sam_both.sh: nonkey_bias is
+# pinned at 0 (amplify_only -- boost key regions only, no suppression) AND the
+# grid was searched with the auction mechanism OFF (see
+# run_ablation_grid_worker.sh), matching the final method exactly.
+# run_eval_attnbias_sam_both_suppress_only.sh has been dropped from
+# run_all_attnbias_sam.sh's comparison set (no longer a live ablation arm), so
+# there's no risk of this sync corrupting it.
+set -e
+
+source /home/jinhua/miniconda3/etc/profile.d/conda.sh
+conda activate huatuovision
+
+LAYER=17
+# must stay identical to run_ablation_grid_worker.sh's copy of these arrays.
+STRENGTHS_TEXT=(1.0 2.0 3.0 4.0 5.0)
+STRENGTHS_VIS=(1.0 2.0 3.0 4.0 5.0)
+
+# ── pick the best (s_text, s_vis) across the full grid, layer fixed at 17.
+#    Each combo's output dir holds exactly one *_result.json (eval_attnbias_sam.py's hparam_tag
+#    already makes the filename unique per layer/strength/auction config, so
+#    no risk of stale files from a previous combo lingering in the same dir),
+#    so glob for it directly instead of hand-reconstructing the filename --
+#    that reconstruction is what broke last time hparam_tag's format changed.
+BEST_S_TEXT=""
+BEST_S_VIS=""
+BEST_SCORE=""
+for S_TEXT in "${STRENGTHS_TEXT[@]}"; do
+    for S_VIS in "${STRENGTHS_VIS[@]}"; do
+        OUTPUT_PATH="results_attnbias_ablation/HuatuoGPT-Vision-7B/sam_both_amplify_noauction_layer${LAYER}_stext${S_TEXT}_svis${S_VIS}"
+        RESULT_F=$(ls "$OUTPUT_PATH"/*_result.json 2>/dev/null | head -n1)
+        if [ -z "$RESULT_F" ]; then
+            echo "[ablation-grid-pick] missing *_result.json under $OUTPUT_PATH -- cannot pick best (s_text, s_vis). Did every run_ablation_grid_worker.sh slice finish?" >&2
+            exit 1
+        fi
+        SCORE=$(python -c "
+import json
+with open('$RESULT_F') as f:
+    d = json.load(f)
+print(d['The total score for multiple-choice questions'])
+")
+        echo "[ablation-grid-pick] s_text=$S_TEXT s_vis=$S_VIS score=$SCORE"
+        if [ -z "$BEST_SCORE" ] || python -c "exit(0 if $SCORE > $BEST_SCORE else 1)"; then
+            BEST_SCORE="$SCORE"
+            BEST_S_TEXT="$S_TEXT"
+            BEST_S_VIS="$S_VIS"
+        fi
+    done
+done
+
+echo "[ablation-grid-pick] best overall: layer=$LAYER s_text=$BEST_S_TEXT s_vis=$BEST_S_VIS (score=$BEST_SCORE)"
+echo "[ablation-grid-pick] syncing into every run_eval_attnbias_sam_*.sh via sync_attn_params.py"
+
+python sync_attn_params.py \
+    --layers_text "$LAYER" --key_bias_text "$BEST_S_TEXT" --nonkey_bias_text "0" \
+    --layers_vis  "$LAYER" --key_bias_vis  "$BEST_S_VIS"  --nonkey_bias_vis  "0"
+
+echo "[ablation-grid-pick] done. All run_eval_attnbias_sam_*.sh now use the ablation-selected layer/strength (nonkey_bias=0, amplify_only style)"
+echo "[ablation-grid-pick] (amplify_only keeps its pinned-zero side; text.sh/vis.sh keep their single-modality side -- see sync_attn_params.py)."

HuatuoGPT-Vision-Abla/results_attnbias_ablation/HuatuoGPT-Vision-7B/sam_both_amplify_noauction_layer17_stext3.0_svis3.0/_snapshot/run_eval_attnbias_sam_both.sh

@@ -0,0 +1,45 @@
+#!/bin/bash
+cd "$(dirname "$0")"
+
+DATA_PATH="/home/jinhua/ExplainableVQA/benchmarks/Medical_Multimodal_Evaluation_Data/medical_multimodel_evaluation_data.json"
+MODEL_PATH="FreedomIntelligence/HuatuoGPT-Vision-7B"
+OUTPUT_PATH="results_attnbias/HuatuoGPT-Vision-7B/sam_both"
+
+SEED=42
+MAX_NEW_TOKENS=256
+
+MEDSAM3_CONFIG="/home/jinhua/ExplainableVQA/HuatuoGPT-Vision-Abla/MedSAM3/configs/full_lora_config.yaml"
+MEDSAM3_DIR="/home/jinhua/ExplainableVQA/HuatuoGPT-Vision-Abla/MedSAM3"
+
+# FINAL METHOD: both mode, amplify-only (nonkey_bias pinned at 0 -- boost key
+# regions, don't suppress anything else) and auction OFF -- every Stage1/SAM-
+# detected key position gets boosted directly, not just an auction-selected
+# subset (ablation found the auction mechanism doesn't help over this simpler
+# design; see run_eval_attnbias_sam_both_noauction.sh, kept around for
+# reference but no longer part of the active comparison set).
+INTERVENTION="both"
+ATTN_LAYERS_TEXT=17
+ATTN_KEY_BIAS_TEXT=5.0
+ATTN_NONKEY_BIAS_TEXT=0.0
+ATTN_LAYERS_VIS=17
+ATTN_KEY_BIAS_VIS=1.0
+ATTN_NONKEY_BIAS_VIS=0.0
+
+MAX_SAMPLES=""
+
+accelerate launch --num_processes=1 eval_attnbias_sam.py \
+    --data_path                "$DATA_PATH"                \
+    --model_path               "$MODEL_PATH"               \
+    --seed                     "$SEED"                     \
+    --max_new_tokens           "$MAX_NEW_TOKENS"           \
+    --output_path              "$OUTPUT_PATH"              \
+    --intervention              "$INTERVENTION"             \
+    --attn_layers_text          "$ATTN_LAYERS_TEXT"         \
+    --attn_key_bias_text        "$ATTN_KEY_BIAS_TEXT"       \
+    --attn_nonkey_bias_text     "$ATTN_NONKEY_BIAS_TEXT"    \
+    --attn_layers_vis           "$ATTN_LAYERS_VIS"          \
+    --attn_key_bias_vis         "$ATTN_KEY_BIAS_VIS"        \
+    --attn_nonkey_bias_vis      "$ATTN_NONKEY_BIAS_VIS"     \
+    --medsam3_config            "$MEDSAM3_CONFIG"           \
+    --medsam3_dir               "$MEDSAM3_DIR"              \
+    ${MAX_SAMPLES:+--max_samples "$MAX_SAMPLES"}