SlowGuess/ABForge-Qwen3-8B-Task2

TEXT GENERATIONPricing:Input $0.468 / Output $1.82Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jun 11, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

SlowGuess/ABForge-Qwen3-8B-Task2 is an 8 billion parameter Qwen3-based model developed by SlowGuess, specifically post-trained for generating detailed ablation experiment design plans. This model excels at producing objectives, setups, variants, fixed protocols, and metrics for ablation studies given a paper's context and a goal. It was fine-tuned using the ABForge pipeline, combining supervised fine-tuning (SFT) and rubric-guided GRPO for high-quality, paper-grounded ablation design.

Loading preview...

Overview

SlowGuess/ABForge-Qwen3-8B-Task2 is an 8 billion parameter model based on the Qwen3 architecture, developed by SlowGuess. It is a specialized model within the ABForge framework, which focuses on post-training pipelines for paper-grounded ablation design. This particular checkpoint has undergone the full ABForge pipeline, involving supervised fine-tuning (SFT) from the base Qwen/Qwen3-8B model, followed by rubric-guided GRPO (Generative Reinforcement Learning with Policy Optimization).

Key Capabilities

  • Ablation Plan Generation: The model's primary function is to generate detailed, controlled ablation experiment design plans. This includes defining the objective, experimental setup, variants, fixed protocols, and metrics.
  • Paper-Grounded Design: It takes a research paper's context and a specific goal as input to produce relevant and coherent ablation plans.
  • Specialized Training: Trained on a dataset derived from CC-licensed research papers, ensuring its outputs are grounded in scientific methodology.

Training and Evaluation

The model was initially SFT on sft_task2_37019.jsonl and subsequently fine-tuned with GRPO on RL_task2_30K.jsonl from the SlowGuess/abforge-data dataset. Evaluation is performed using the held-out AblationBench split (ablationbench_200.jsonl) of the same dataset. Users can reproduce the AblationBench evaluation using the provided SlowGuess/Abforge_1 code.

Good For

  • Researchers and scientists needing assistance in designing rigorous ablation studies.
  • Automating the generation of experimental plans for evaluating specific components or hypotheses within a research paper's context.
  • Applications requiring structured, detailed outputs for experimental design.

Finetunes
3 models