[doc] FAQ

[tpoisonooo]

New access links

Try it here:

• https://modelscope.cn/studios/chenzihong/GraphGen
• https://huggingface.co/spaces/chenzihong/GraphGen

[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

[tpoisonooo]

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.

[tpoisonooo]

Why all of output loss are same value?

See #17 and try vllm serve with v0.6.6post1 version.

logprobs in v0.8.3 still fails vllm-project/vllm#16838

[unlimited0715]

What should I fill in for the tokenizer if I use a large language model deployed locally via Ollama as the Synthesizer Model?

[ChenZiHong-Gavin]

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".