SlowGuess/ABForge-Qwen3-8B-Task1

ABForge-Qwen3-8B-Task1: Ablation Objective Identification

This model, developed by SlowGuess, is an 8 billion parameter variant of the Qwen3-8B architecture, specifically engineered for Task 1: Ablation Objective Identification within the ABForge framework. ABForge is a post-training pipeline focused on paper-grounded ablation design.

Key Capabilities

• Proposes Ablation Objectives: Given the ablation-free context of a research paper, the model identifies and suggests potential ablation objectives.
• Structured Output: Each proposed objective is presented as a pair: a Target Module (the component to be ablated) and a Research Question that the ablation aims to answer.
• Specialized Training: The model undergoes a unique post-training pipeline, starting with supervised fine-tuning (SFT) from Qwen/Qwen3-8B, followed by rubric-guided GRPO (SFT → GRPO).
• Data Grounding: Training data is derived from CC-licensed research papers, specifically from the SlowGuess/abforge-data dataset, ensuring relevance to academic research.

When to Use This Model

This model is ideal for researchers and developers working on:

• Automated Research Analysis: Identifying critical components and their associated research questions within scientific papers.
• Ablation Study Design: Assisting in the systematic design of ablation experiments by suggesting relevant modules and hypotheses.
• Understanding Research Methodology: Gaining deeper insights into the experimental design and component importance in published works.

Evaluation of this model's performance on the held-out AblationBench split can be reproduced using the provided SlowGuess/Abforge_1 code and scripts.