ABForge-Qwen3-8B

📄 ArXiv  ｜  💻 Code  ｜  🤗 Collection

About

This repository contains ABForge-Qwen3-8B, the model presented in ABForge: Post-Training for Paper-Grounded Ablation Design. Given a paper's methodology with its ablation content removed, it proposes the ablation objectives the paper should investigate and designs a rigorous experiment plan for each — both from a single checkpoint.

The model is obtained through an SFT → GRPO pipeline on Qwen3-8B, trained throughout on a 1:1 mixture of the two tasks:

1. Supervised fine-tuning on the ABForge SFT corpus, one full epoch → ABForge-Qwen3-8B-SFT
2. Rubric-guided GRPO, 200 updates, each rollout routed to its task's reward by data_source → this model

Links

Resource	Link
Code	SlowGuess/Abforge_1
Training & evaluation data	SlowGuess/abforge-data
SFT checkpoint	SlowGuess/ABForge-Qwen3-8B-SFT
RL-only checkpoint	SlowGuess/ABForge-Qwen3-8B-RL
Per-paper outputs & judge rationales	outputs/task{1,2}/*/abforge.jsonl in the data repo

Training data

SlowGuess/abforge-data ships a single table under train/, one row per paper, built by a semi-automated audit-in-the-loop pipeline over research papers from major ML, NLP and CV venues. Both stages select their rows from that table by its split columns and mix the two tasks 1:1:

• SFT — in_sft_task1 (45,961 papers) + in_sft_task2 (36,955)
• GRPO — in_rl_task1 + in_rl_task2 (30,000 papers each, disjoint from the SFT pool)

The benchmark papers carry no training flag, so they cannot leak into either stage. The preprocessing scripts in the code release turn the table into the training parquets directly.

Performance

AblationBench, automated rubric-based LLM-as-a-Judge evaluation (eval/ablationbench_200.jsonl, 200 papers, judge claude-sonnet-4-6). Task 1 is ablation objective identification (paper_score); Task 2 is ablation plan synthesis (design_score, ×100).

Model	Task 1	Task 2
Qwen3-8B (base)	44.4	43.4
ABForge-Qwen3-8B-SFT (SFT only)	30.7	52.2
ABForge-Qwen3-8B-RL (RL only)	52.2	54.9
ABForge-Qwen3-8B (SFT → GRPO)	55.9	62.4

Each stage on its own falls short: SFT trades Task 1 for Task 2, and GRPO from the base model plateaus below the full pipeline. Running GRPO from the SFT checkpoint is worth +3.7 on Task 1 and +7.5 on Task 2 over RL alone.

Evaluation

Reproduce the numbers above with the code release:

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

python run_inference_local.py --task 1 \
  --input data/eval/ablationbench_200.jsonl \
  --output outputs/task1_infer.jsonl \
  --model-path SlowGuess/ABForge-Qwen3-8B \
  --dtype bf16 --device-map auto \
  --max-new-tokens 5120 --temperature 0.0 --stop-on '</Result>'

export JUDGE_API_BASE=https://api.openai.com/v1
export JUDGE_API_KEY=...
export JUDGE_MODEL=...
scripts/evaluate_task1.sh outputs/task1_infer.jsonl

Swap --task 2, --stop-on '</Proposed_Plan>' and scripts/evaluate_task2.sh for Task 2. The model is trained on the prompt templates in the code release and the rubric evaluator expects the matching output structure, so use those templates and greedy decoding.

Citation

@misc{abforge2026,
  title={ABForge: Post-Training for Paper-Grounded Ablation Design},
  author={TODO},
  year={2026},
}