SlowGuess/ABForge-Qwen3-8B-Task1-SFT
SlowGuess/ABForge-Qwen3-8B-Task1-SFT is an 8 billion parameter, 32K context length model fine-tuned from Qwen/Qwen3-8B. It is specifically designed for the ABForge post-training pipeline, focusing on Task 1: Ablation Objective Generation. This model excels at proposing candidate ablation objectives, comprising a Target Module and a Research Question, based on the ablation-free context of a research paper.
Loading preview...
Model Overview
SlowGuess/ABForge-Qwen3-8B-Task1-SFT is an 8 billion parameter language model, supervised fine-tuned (SFT) from the Qwen/Qwen3-8B base model. It is a core component of the ABForge post-training pipeline, which aims to facilitate paper-grounded ablation design in research.
Key Capabilities
- Ablation Objective Generation: The primary function of this model is to generate candidate ablation objectives. Given the ablation-free context of a research paper, it proposes a pair consisting of a Target Module (the component to ablate) and a Research Question that the ablation is intended to answer.
- Research Paper Analysis: It processes and understands the context of research papers to identify potential areas for ablation studies.
Training and Evaluation
The model was SFT on the train/sft_task1_45961.jsonl dataset, derived from CC-licensed research papers, available in the SlowGuess/abforge-data dataset. Evaluation is performed using the held-out AblationBench split (eval/ablationbench_200.jsonl) from the same dataset. Users can reproduce the AblationBench evaluation using the provided SlowGuess/Abforge_1 code.
Related Models
This model is part of a series of ABForge models for Task 1, including SlowGuess/ABForge-Qwen3-8B-Task1 and SlowGuess/ABForge-Qwen3-8B-Task1-RL.