README.md

GraphFRA-X-DPO

Graph-conditioned DPO checkpoint initialized from dharun2049/GraphFRA-X-SFT.

Preference source: HKUST-DSAIL/Graph-R1-dataset-complete (MIT).

Configuration:

• beta: 0.1
• loops: 2
• length normalization: False
• graph encoder LR: 5e-06
• decoder LR: 2e-05
• bridge LR: 5e-05
• preference pairs: 36000

Files:

• graph_encoder.pt
• graph_bridge.pt
• graph_config.json
• decoder/
• dpo_config.json