Modelpublic

Mintzs/ora-qwen3.5-4b-docsynth-v1

sourceHugging Faceupdated 25d agoView on Hugging Face

0likes21downloads

Model Card

Ora Qwen3.5-4B Document Synthesis V1

Research-only document synthesis checkpoint trained with completion-only, language-model LoRA while preserving Qwen3.5's multimodal vision tower.

—Base revision: 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a
—Dataset release: ora_document_synthesis_research_v1 v1.2.0
—Dataset digest: d35206e2fe61196d85db02027a148015eac4b8e9d47d0287d1cb9dc49b5878b1
—Training: one epoch, rank-32 LoRA, BF16 base, LR 4e-05
—Root weights: merged BF16 SafeTensors
—adapter/: unmerged LoRA adapter
—bnb-int8/: standalone bitsandbytes INT8 Hugging Face checkpoint

The mixed upstream dataset is licensed for Ora research use only. Benchmark the checkpoint independently before drawing capability claims.