#!/usr/bin/env sh
# Auto-generated combo: model=Qwen3-4B source=tri config=prompt_based_separator (pbs)
# head ranking fixed to all_roc_inst_0.1; TOPK default is overridable via env.
#
# In-training quick eval uses a DEV HOLDOUT of this run's own training source
# (tri): the last DEV_HOLDOUT source rows are dropped from training and used
# only for the training_log.csv rows. It never touches the 4-1 cross-eval set.
set -eu

SCRIPT_DIR=$(CDPATH= cd -- "$(dirname -- "$0")" && pwd)
TRAIN_DIR=$(CDPATH= cd -- "$SCRIPT_DIR/.." && pwd)
BASE=$(CDPATH= cd -- "$TRAIN_DIR/.." && pwd)
ROOT=$(CDPATH= cd -- "$BASE/.." && pwd)

export MODEL_PATH="${MODEL_PATH:-$ROOT/models/Qwen3-4B}"
export DATA_PATH="${DATA_PATH:-$BASE/3-1_model_training_data_gen/single_turn/tri_prompt_based_separator.json}"
export HEAD_PATH="${HEAD_PATH:-$BASE/2-2_head_identification_scoring/model_score/sep_Qwen3-4B/heads_sorted/all_roc_inst_0.1.json}"
export OUTPUT_DIR="${OUTPUT_DIR:-$TRAIN_DIR/outputs_lora/qwen3-4b_tri_pbs_all_roc_inst_0.1}"
export TOPK="${TOPK:-12.5p}"

# quick-eval dev slice: same source as training, config matches, held-out tail rows
export EVAL_DATA_PATH="${EVAL_DATA_PATH:-$BASE/1_raw_dataset/topicattack/data/crafted_instruction_data_tri_injection_qa.json}"
export EVAL_TOPIC_PATH="${EVAL_TOPIC_PATH:-$BASE/1_raw_dataset/topicattack/data/crafted_instruction_data_tri_conversation_attack_complete.json}"
export EVAL_CONFIG="${EVAL_CONFIG:-prompt_based_separator}"
export DEV_HOLDOUT="${DEV_HOLDOUT:-24}"

exec "$TRAIN_DIR/train_attn_kl_clean.sh"