GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
SlowGuess
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
SlowGuess
Model Tree
Qwen/Qwen3-8B
Input Modalities
Output Modalities
Supported Functionality
Both ABForge tasks in one checkpoint — Task 1: Ablation Objective Generation (propose Target Module / Research Question pairs for a paper) and Task 2: Ablation Experiment Design (produce an executable ablation plan).
Mixed-task SFT on train/sft_task1_45961.jsonl + train/sft_task2_37019.jsonl from
SlowGuess/abforge-data
(derived from CC-licensed research papers), shuffled 1:1, one full epoch.
Full AblationBench, automated rubric-based LLM-as-a-Judge evaluation:
| Stage | Task 1 (%) | Task 2 (%) |
|---|---|---|
Qwen/Qwen3-8B (base) | 44.4 | 43.4 |
| This model (unified, SFT only) | 30.7 | 52.2 |
-Combined (unified, SFT → GRPO) | 55.9 | 62.4 |
SFT alone improves Task 2 but degrades Task 1 relative to the base model; it serves as an effective initialization for the subsequent joint RL stage rather than as a final model.
SlowGuess/ABForge-Qwen3-8B-Combined — the released unified modelSlowGuess/ABForge-Qwen3-8B-Task1-SFT / -Task2-SFT — task-specific SFT counterpartsReproduce AblationBench evaluation with the
SlowGuess/Abforge_1 code.