Repository navigation
[doc] FAQ #10
Description
Activity
Why not CP (Continue Pretraining) / How about base model know nothing about the domain knowledge ?
When SFT data is sufficient, the model can also learn new knowledge. Completing "data cleaning-CP-SFT-DPO" from scratch costs hundreds of thousands of dollars. As an engineer rather than an academic researcher, GraphGen is the best solution we can give.
- pinned this issue
on Apr 27, 2025 Why all of output loss are same value?
See #17 and try
vllm servewithv0.6.6post1version.logprobsinv0.8.3still fails vllm-project/vllm#16838What should I fill in for the tokenizer if I use a large language model deployed locally via Ollama as the Synthesizer Model?
What should I fill in for the tokenizer if I use a large language model deployed locally via Ollama as the Synthesizer Model?
You can use the default tiktoken tokenizers like "cl100k_base".
New access links
Try it here:
[Deprecated] Cannot download result on OpenXLab
This is https vs http problem. Just use https://g-app-center-120612-6433-jpdvmvp.openxlab.space instead of https://openxlab.org.cn/apps/detail/chenzihonga/GraphGen
Or you can:
STEP1. right-click download button and "copy link"
STEP2. open a new broswer tab, input the link