BREAKING
GraphGen Released for LLM Data
How GraphGen Works
1
Build KG from text
↓
2
Find gaps via ECE
↓
3
k-hop sampling
↓
4
Generate QA data
SFT Gains on Qwen2.5-7B
SeedBench base
51.5
SeedBench +GG
65.9
AIME25 base
7.2
AIME25 +GG
22.7
0
LawBench
0
MedQA
0
pt
Pretrain avg gain
Strengths and Challenges
Strengths
●
Fills long-tail gaps
●
Reduces hallucination
●
Broad input support
Challenges
●
KG build cost
●
Noise robustness
●
Domain adaptation
KG Meets LLM Training
AI NEWS BLITZ
An open-source framework builds LLM training data using knowledge graphs.