G-reasoner Configs

This page groups the task presets under gfmrag/workflow/config/gfm_reasoner/.

Directory Layout

File	Purpose	Typical entrypoint
index_dataset.yaml	Build processed/stage1/ from raw data	python -m gfmrag.workflow.index_dataset --config-name gfm_reasoner/index_dataset
kgc_trianing.yaml	Run KGC pretraining for the G-reasoner model family	python -m gfmrag.workflow.kgc_training --config-name gfm_reasoner/kgc_trianing
qa_inference.yaml	Run QA from saved retrieval outputs	python -m gfmrag.workflow.qa --config-name gfm_reasoner/qa_inference
sft_training.yaml	Run supervised fine-tuning	python -m gfmrag.workflow.sft_training --config-name gfm_reasoner/sft_training
sft_training_w_answer.yaml	Run SFT with additional answer supervision	python -m gfmrag.workflow.sft_training --config-name gfm_reasoner/sft_training_w_answer
stage3_qa_ircot_inference.yaml	Run retrieval plus IRCOT-style reasoning	python -m gfmrag.workflow.qa_ircot_inference --config-name gfm_reasoner/stage3_qa_ircot_inference
visualize_path.yaml	Visualize reasoning paths on dataset examples	visualization workflow

How To Read This Folder

The gfm_reasoner presets follow a stable pattern:

• hydra.run.dir controls the output root.
• defaults pulls in shared component groups such as ner_model, openie_model, el_model, text_emb_model, and wandb.
• task-specific sections such as dataset, datasets, graph_retriever, model, trainer, llm, and test then override the shared pieces.

Main Pages In This Docs Section

• G-reasoner Graph Index Config
• G-reasoner Retrieval and QA Config
• G-reasoner SFT Training Config

Shared Component Pages Used By This Folder

• NER Model Config
• OpenIE Model Config
• Entity Linking Model Config
• Text Embedding Config
• Wandb Config