SkyworkAI / SkyworkAI/Skywork-R1V
关于R1系列<模块重组>的几点疑问
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 3.2k
- Forks
- 282
- PR merge metrics
- No merged PRs in 30d
Description
skywork团队你好,非常感谢能看到贵团队富有创造力的工作,有几点疑问想请教一下:
-
V2技术报告3.1提到的模块重组,这里是否进行了mlp module的对齐训练?还是说如报告中提到的'eliminate SFT stage',完全摒弃了任何形式的监督训练,以一个随机初始化的projector来进行后续mpo+grpo呢?
-
第一个问题的延续,V1报告中披露了三阶段对齐训练使用了2M数据,这是一个远远小于主流VLM在对齐阶段的数据量,且重组后能达到60.2的MMMU分数(比llava-ov-72b公布的分数高,数据量仅有它的1/4)。我好奇的是:1) 这个收益是源自对齐训练本身还是源自从32b-ins到32b-distill的重组?2) 其他通用视觉理解、感知任务的性能有观察么?例如chart/ocr等。因为有可能这种重组会导致在通用任务上出现很大的视觉幻觉 3) 不知道是否方便透露2M数据的大致构成,从分数上看,这个数据的训练效率比onevision的要高很多
最后还是感谢贵组的工作提供了很多研究探索思路!
Contributor guide
No contributing guide indexed for this repository
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start with the V2 technical report section 3.1 and the V1 report's description of the three-stage alignment training. Trace the module reorganization, alignment stages, 2M-data composition, and reported chart/OCR or general vision results; done means documenting clear answers to each question.
Written by the indexing model from the issue text.
Assessment
- Domain
- ai, documentation, machine-learning
- Issue type
- Documentation
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 20/100