SlowGuess/ABForge-Qwen3-8B-Task1-RL

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

ABForge-Qwen3-8B-Task1-RL is an 8 billion parameter Qwen3-based language model developed by SlowGuess, specifically fine-tuned for identifying ablation objectives in research papers. This model proposes candidate ablation objectives, consisting of a Target Module and a Research Question, given the ablation-free context of a paper. It was trained using GRPO directly from Qwen/Qwen3-8B, optimizing a fixed rubric-based reward, making it specialized for structured scientific text analysis.

Loading preview...

Overview

ABForge-Qwen3-8B-Task1-RL is an 8 billion parameter model based on the Qwen3 architecture, developed by SlowGuess. It is a specialized model within the ABForge post-training pipeline, designed for Task 1: Ablation Objective Identification. This model was trained using GRPO (Gradient-based Reward Policy Optimization) directly from the base Qwen/Qwen3-8B model, without a supervised warm-start, and optimizes a fixed rubric-based reward.

Key Capabilities

  • Ablation Objective Identification: Given the ablation-free context of a research paper, the model proposes candidate ablation objectives.
  • Structured Output: Each proposed objective is expressed as a pair: a Target Module (the component to ablate) and a Research Question it aims to answer.
  • GRPO Training: Utilizes GRPO on the RL_task1_30K.jsonl dataset from SlowGuess/abforge-data, which is derived from CC-licensed research papers.

Good for

  • Scientific Text Analysis: Researchers and developers working with academic papers to automatically identify potential ablation studies.
  • Automated Experiment Design: Generating structured suggestions for experimental design based on existing research contexts.
  • Understanding Research Paper Structure: Extracting key components and research questions related to experimental methodology.

Evaluation of this model can be reproduced using the SlowGuess/Abforge_1 code, specifically on the held-out AblationBench split (ablationbench_200.jsonl) of the training dataset.