SlowGuess/ABForge-Qwen3-8B
ABForge-Qwen3-8B is an 8 billion parameter language model developed by SlowGuess, based on the Qwen3-8B architecture, with a 32768 token context length. It is specifically fine-tuned for paper-grounded ablation design, excelling at proposing ablation objectives and designing rigorous experiment plans from research paper methodologies. This model is optimized through a Supervised Fine-Tuning (SFT) and Rubric-Guided Reinforcement Learning from Human Feedback (GRPO) pipeline, making it highly effective for academic research tasks involving experimental design.
Loading preview...
Overview
ABForge-Qwen3-8B is an 8 billion parameter model developed by SlowGuess, derived from the Qwen3-8B base model. It is specialized in paper-grounded ablation design, capable of proposing ablation objectives and designing experimental plans based on research paper methodologies where ablation content has been removed.
Key Capabilities
- Ablation Objective Identification: Identifies relevant ablation objectives from research paper methodologies.
- Ablation Plan Synthesis: Designs rigorous experimental plans for identified ablation objectives.
- Specialized Training: Achieved through a unique SFT → GRPO (Supervised Fine-Tuning followed by Rubric-Guided Reinforcement Learning from Human Feedback) pipeline, trained on a 1:1 mixture of these two tasks.
- Performance: Outperforms its base model (Qwen3-8B) and SFT-only/RL-only checkpoints on the AblationBench evaluation, scoring 55.9 on Task 1 (objective identification) and 62.4 on Task 2 (plan synthesis).
Good For
- Academic Research: Assisting researchers in designing and refining ablation studies for their papers.
- Methodology Analysis: Analyzing and understanding the experimental design aspects of scientific literature.
- Automated Experiment Design: Generating structured experimental plans based on textual descriptions of research methodologies.