# Abforge_Training

Local training workspace for ABForge-style reinforcement learning on scientific ablation design.

## Scope

This repository contains:

- `verl_proj/`: a customized `verl` training tree
- `reward_part/`: local LLM-as-judge services for Task 1 and Task 2 rewards
- `data/`: local raw RL data
- `ARR_March___Ablation_Study_Post_Training__New_-14.pdf`: reference paper

The current setup focuses on Qwen3-based GRPO training for:

- Task 1: ablation objective identification
- Task 2: ablation experiment synthesis

## Added Local Components

### Task 2

- Data preprocess: `verl_proj/examples/data_preprocess/abforge_task2_rl.py`
- Judge service: `reward_part/task2_call_api.py`
- Reward API: `reward_part/task2_rw.py`
- Training script: `verl_proj/examples/grpo_trainer/run_abforge_qwen3_8b_hf.sh`

### Task 1

- Data preprocess: `verl_proj/examples/data_preprocess/abforge_task1_rl.py`
- Judge service: `reward_part/task1_call_api.py`
- Reward API: `reward_part/task1_rw.py`
- Training script: `verl_proj/examples/grpo_trainer/run_abforge_qwen3_8b_task1_hf.sh`

## Notes

- Reward is computed only on the structured output block:
  - Task 1: `<Result>...</Result>`
  - Task 2: `<Proposed_Plan>...</Proposed_Plan>`
- Qwen3 training paths should disable `enable_thinking` in chat templating.
- The current training scripts use `hf` rollout first for compatibility bring-up. vLLM can be added later after environment validation.