ABForge-Qwen3-8B-Task2-RL

An ABForge model for Task 2: Ablation Plan Generation.

ABForge is a post-training pipeline for paper-grounded ablation design. This checkpoint is trained with GRPO directly from Qwen/Qwen3-8B (no supervised warm-start), optimizing a fixed rubric-based reward (GRPO).

Task

Given a paper's context and a goal, the model produces a detailed, controlled ablation experiment design plan (objective, setup, variants, fixed protocols and metrics).

Training data

GRPO on train/RL_task2_30K.jsonl, from SlowGuess/abforge-data (derived from CC-licensed research papers). Evaluation uses the held-out AblationBench split (eval/ablationbench_200.jsonl) of the same dataset.

Related models (Task 2)

• SlowGuess/ABForge-Qwen3-8B-Task2
• SlowGuess/ABForge-Qwen3-8B-Task2-SFT
• SlowGuess/ABForge-Qwen3-8B-Task2-RL (this model)

Evaluation

Reproduce AblationBench evaluation with the SlowGuess/Abforge_1 code:

git clone https://github.com/SlowGuess/Abforge_1 && cd Abforge_1
huggingface-cli download SlowGuess/abforge-data --repo-type dataset --local-dir data

export MODEL_PATH=SlowGuess/ABForge-Qwen3-8B-Task2-RL

# 1. Generate predictions on AblationBench (writes scorer-ready JSONL)
python run_inference_local.py --task 2 \
  --input data/eval/ablationbench_200.jsonl \
  --output preds.jsonl \
  --model-path "$MODEL_PATH" --dtype bf16 --device-map auto \
  --max-new-tokens 4096 --temperature 0.0 --stop-on '</Proposed_Plan>'

# 2. Score against the fixed 10-item AblationBench rubric
export JUDGE_API_BASE=... JUDGE_API_KEY=... JUDGE_MODEL=...
scripts/evaluate_task2.sh preds.jsonl

Links

• Dataset: SlowGuess/abforge-data
• Code: SlowGuess/Abforge_1

Citation

@misc{abforge,
  title  = {ABForge: A Post-Training Pipeline for Paper-Grounded Ablation Design},
  author = {ABForge authors},
  year   = {2026},
  howpublished = {\url{https://github.com/SlowGuess/Abforge_1}}
}