SlowGuess/ABForge-Qwen3-8B-RL

Overview

SlowGuess/ABForge-Qwen3-8B-RL is an 8 billion parameter model derived from Qwen3-8B, specifically trained using rubric-guided GRPO (Generative Reinforcement Learning with Policy Optimization) for 200 updates. This model represents the "RL-only ablation" of the ABForge project, focusing on post-training for paper-grounded ablation design without an initial Supervised Fine-Tuning (SFT) stage. It is designed to analyze research paper methodologies, propose relevant ablation objectives, and formulate detailed experimental plans.

Key Capabilities

• Ablation Objective Identification: Excels at identifying potential ablation objectives from a given paper's methodology.
• Ablation Plan Synthesis: Capable of designing rigorous experimental plans for the identified ablation objectives.
• Specialized Training: Trained directly on Qwen3-8B with GRPO, using a 1:1 mixture of two tasks related to ablation design.
• Performance Improvement: Achieves a Task 1 score of 52.2 and a Task 2 score of 54.9 on the AblationBench, outperforming the base Qwen3-8B model in both tasks.

When to Use This Model

This model is particularly suited for research and development workflows that involve:

• Automated Scientific Analysis: Generating ablation objectives and experimental designs from research papers.
• Methodology Auditing: Assisting in the critical review and extension of scientific methodologies.
• Academic Research Tools: Developing tools that aid researchers in structuring their experimental work and identifying gaps in existing studies.

It is important to note that while this model shows strong performance in its specialized tasks, the full ABForge-Qwen3-8B model (which includes an SFT warm-start) generally achieves higher scores (55.9 for Task 1 and 62.4 for Task 2), indicating that SFT can serve as an effective initialization for the RL stage.