Training and Inference Scripts

This directory contains GitHub-ready versions of the local training and inference scripts. Private paths and API keys have been removed. Configure paths through environment variables.

Files

• sft.sh: supervised fine-tuning with swift sft.
• RL.sh: GRPO/RLHF training with an external rollout server.
• plugin_reward.py: custom reward functions used by RL.sh.
• 1.infer_task_no_prompt.py: task-level inference without injecting the prompt tags.
• 1.infer_task_prompt_tags.py: task-level inference with the built-in GeoGebra prompt.
• .env.example: example configuration values. Do not commit real API keys.

Quick Start

cp .env.example .env
# Edit .env, then:
source .env

Run SFT:

OUTPUT_DIR=./outputs/sft bash sft.sh

Run RL training:

OUTPUT_DIR=./outputs/rl_both bash RL.sh

Run evaluation:

CHECKPOINT=/path/to/checkpoint TASK_DATASET_DIR=./task_datasets python 1.infer_task_prompt_tags.py
CHECKPOINT=/path/to/checkpoint TASK_DATASET_DIR=./task_datasets WORLD_SIZE=8 python 1.infer_task_no_prompt.py

Notes

• SWANLAB_API_KEY is intentionally not set in the scripts. Export it locally only when needed.
• REWARD_PLUGIN, SYSTEM_PROMPT_FILE, model checkpoints, and datasets are expected to be provided by the user or the surrounding project.