xiaolesu/OsmosisProofling-SFT
TEXT GENERATIONConcurrency Cost:1Model Size:8BQuant:FP8Ctx Length:32kPublished:Mar 24, 2026Architecture:Transformer0.0K Cold
OsmosisProofling-SFT by xiaolesu is an experimental SFT-only checkpoint based on Qwen3-8B, developed for autoformalization research. It was trained on 20,000 heterogeneous natural-language and Lean 4 pairs, with thinking disabled. This model is specifically designed to explore data overlap as a post-training hyperparameter for autoformalization tasks, as detailed in its associated research paper.
Loading preview...
OsmosisProofling-SFT Overview
xiaolesu/OsmosisProofling-SFT is an experimental SFT-only checkpoint derived from the Qwen3-8B architecture. This model is a key artifact from the research paper "Data Overlap as a Post-Training Hyperparameter for Autoformalization" by Xiaole Su, Kasey Zhang, and Andy Lyu.
Key Characteristics
- Base Model: Qwen3-8B, with its 'thinking' mechanism disabled for this specific variant.
- Training Data: Fine-tuned on a dataset comprising 20,000 heterogeneous pairs, which include both natural language and Lean 4 formalization examples.
- Research Focus: Primarily developed to investigate the impact of data overlap as a hyperparameter in the context of autoformalization, aiming to translate natural language into formal mathematical proofs (Lean 4).
- Experimental Nature: This model represents an intermediate SFT-only stage of the research, with further details and results available in the paper repository.
Good for
- Autoformalization Research: Ideal for researchers and practitioners exploring methods for converting natural language into formal mathematical proofs, particularly within the Lean 4 ecosystem.
- Understanding Data Overlap: Useful for studying the effects of data overlap in supervised fine-tuning for complex reasoning tasks.
- Benchmarking: Can serve as a baseline or comparison point for future autoformalization models, especially those based on Qwen3-8B or similar architectures.