1.5

Release v1.5

This is the first official release of LLaVA-OneVision-1.5, focusing on documentation improvements, data packing & preprocessing reliability, training configuration defaults, evaluation additions, and multiple bug fixes.

Highlights

• Improved documentation and onboarding experience (README, training TODOs, LICENSE updates)
• More robust data packing / WebDataset conversion with support for mixed packing
• Added SFT data preprocessing guidance and materials
• Introduced evaluation-related content
• Improved training defaults (e.g., dtype bfloat16, default 4B merge, auto config updates)
• Numerous bug fixes across packing, filtering, and demo workflows

---

What’s Changed

Documentation & Project Maintenance

• Update README by @yiyexy in #1
• Add training framework todo list in Readme.md by @VVsssssk in #10
• docs: update LICENSE by @RobitYadda in #21
• offline_sample_packing description by @fdcp in #34
• Update README.md by @wkzhang636 in #77
• Delete run.sh by @anxiangsir in #7

Data Processing / Packing / Filtering

• fix packing bugs by @chengzheng345 in #3
• Update the convert_to_webdataset.py script to support mixed data pack… by @yiyexy in #4
• Fix some bugs by @yiyexy in #5
• add data filter by @chengzheng345 in #12
• fix cfg['filter_with_caption'] bug by @chengzheng345 in #27
• bug fix : processing samples whose names contain some ':' by @fdcp in #49

Training Defaults & Configuration

• set processor.image_processor.max_pixels = 1600*1600 by @yiyexy in #8
• default merge 4b by @yiyexy in #13
• auto update config by @yiyexy in #15
• update default dtype to bfloat16 by @yiyexy in #18
• Remove the hosts Settings for multiple machines and only support sing… by @chengzheng345 in #56

Evaluation

• add Evaluation by @mathCrazyy in #16

SFT Data

• Add SFT data preprocess by @wideyard in #22
• Update sft data md & new image by @wideyard in #25

Bug Fixes & Robustness Improvements

• fix a bug by @yiyexy in #30
• fix_issue#31 by @killTheHostage in #33
• Remove duplicate dependency py-cpuinfo libraries. by @Lornatang in #50
• Fixed the issue that the model_path name does not correspond to the HF, causing the model to fail to load. by @Lornatang in #51
• refactor(merge_model): Enhanced file implementation robustness by @Lornatang in #52
• fix Stage 1.5 Mid-Training demo error. by @Lornatang in #57

---

New Contributors

• @yiyexy made their first contribution in #1
• @chengzheng345 made their first contribution in #3
• @anxiangsir made their first contribution in #7
• @VVsssssk made their first contribution in #10
• @mathCrazyy made their first contribution in #16
• @RobitYadda made their first contribution in #21
• @wideyard made their first contribution in #22
• @killTheHostage made their first contribution in #33
• @fdcp made their first contribution in #34
• @Lornatang made their first contribution in #50
• @wkzhang636 made their first contribution in #77

Full Changelog: https://github.com/EvolvingLMMs-Lab/LLaVA-OneVision-1.5/commits/1.5