agentic-rl-main / scripts /test /train_opd.sh
Jack04810's picture
Add files using upload-large-folder tool
d13981b verified
Raw History Blame Contribute Delete
854 Bytes
#!/usr/bin/env bash
# Fast baseline: cross-model OPD (7B teacher + 0.5B student) with DeepSpeed.
# DyME-aligned routing (teacher-probe gated OPD, no embedded cold-start, no Visual Supervision).
set -euo pipefail
TEST_DIR="$(cd "$(dirname "$0")" && pwd)"
source "${TEST_DIR}/launch_utils.sh"
CONFIG_PATH="config/config_opd_7b_deepspeed.yaml"
export ACCELERATE_CONFIG="${ACCELERATE_CONFIG:-$(resolve_deepspeed_zero1_config)}"
prepare_fast_test_data "${CONFIG_PATH}"
ensure_spacy_model
NUM_PROCESSES="$(detect_num_gpus)"
print_fast_plan "opd" "${CONFIG_PATH}"
LOG_FILE="$(fast_train_log_path train_test_opd)"
run_train_with_log "${LOG_FILE}" \
"${PYTHON_BIN}" -m accelerate.commands.launch --config_file "${ACCELERATE_CONFIG}" --num_processes "${NUM_PROCESSES}" main.py \
--config "${CONFIG_PATH}" \
--mode rl \
--opsd_enabled \
"$@"