SFT Constructor Configuration¶

SFT constructors turn raw QA files into processed supervision data for downstream training and evaluation.

GFM-RAG SFT Constructor¶

An example GFM-RAG SFT constructor configuration file is shown below:

gfm_rag_sft_constructor

gfmrag/workflow/config/sft_constructor/gfm_rag_sft_constructor.yaml

_target_: gfmrag.graph_index_construction.sft_constructors.GFMRAGConstructor # The SFTConstructor class
root: tmp/qa_construction # Temporary directory for storing intermediate files during SFT construction
ner_model: ${ner_model}
el_model: ${el_model}
num_processes: 10 # Number of processes to use
force: False # Whether to force recompute the QA data

Parameter	Options	Note
`_target_`	`gfmrag.graph_index_construction.sft_constructors.GFMRAGConstructor`	The class name of `GFMRAGConstructor`.
`root`	`tmp/qa_construction`	Temporary directory for processed intermediate files.
`ner_model`	`${ner_model}`	NER config used to identify start entities.
`el_model`	`${el_model}`	EL config used to map entities into the graph.
`num_processes`	Positive integer	Number of worker processes used during preprocessing.
`force`	`True`, `False`	Whether to recompute processed outputs.

GFM-Reasoner SFT Constructor¶

An example GFM-Reasoner SFT constructor configuration file is shown below:

gfm_reasoner_sft_constructor

gfmrag/workflow/config/sft_constructor/gfm_reasoner_sft_constructor.yaml

_target_: gfmrag.graph_index_construction.sft_constructors.GFMReasonerConstructor # The SFTConstructor class
root: tmp/qa_construction # Temporary directory for storing intermediate files during SFT construction
ner_model: ${ner_model}
el_model: ${el_model}
num_processes: 10 # Number of processes to use
force: False # Whether to force recompute the QA data

Parameter	Options	Note
`_target_`	`gfmrag.graph_index_construction.sft_constructors.GFMReasonerConstructor`	The class name of `GFMReasonerConstructor`.
`root`	`tmp/qa_construction`	Temporary directory for processed intermediate files.
`ner_model`	`${ner_model}`	NER config used to identify start entities.
`el_model`	`${el_model}`	EL config used to map entities into the graph.
`num_processes`	Positive integer	Number of worker processes used during preprocessing.
`force`	`True`, `False`	Whether to recompute processed outputs.

HippoRAG2 SFT Constructor¶

An example HippoRAG2 SFT constructor configuration file is shown below:

hipporag2_sft_constructor

gfmrag/workflow/config/sft_constructor/hipporag2_sft_constructor.yaml

_target_: gfmrag.graph_index_construction.sft_constructors.HippoRAG2Constructor # The SFTConstructor class
root: tmp/qa_construction # Temporary directory for storing intermediate files during SFT construction
text_emb_model: ${text_emb_model}
enable_filtering: True # Whether to enable filtering of the constructed QA data using a language model
num_processes: 1 # Number of processes to use
topk: 5 # Top-k nodes to be selected for each question for SFT data construction
llm_for_filtering: gpt-4o-mini # The name of the language model to use for filtering facts
retry: 5 # Number of retries for LLM inference in case of failure
force: False # Whether to force recompute the QA data
start_type: ["entity", "document"]
target_type: ["entity", "document"]

Parameter	Options	Note
`_target_`	`gfmrag.graph_index_construction.sft_constructors.HippoRAG2Constructor`	The class name of `HippoRAG2Constructor`.
`root`	`tmp/qa_construction`	Temporary directory for processed intermediate files.
`text_emb_model`	`${text_emb_model}`	Text embedding model used for candidate generation.
`enable_filtering`	`True`, `False`	Whether to run LLM-based filtering over candidate facts.
`num_processes`	Positive integer	Number of worker processes.
`topk`	Positive integer	Number of candidate nodes selected per question.
`llm_for_filtering`	Model name	LLM used for fact filtering (e.g. `gpt-4o-mini`).
`retry`	Positive integer	Retry count for filtering calls.
`force`	`True`, `False`	Whether to recompute processed outputs.
`start_type`	List of node types	Node types allowed in start nodes.
`target_type`	List of node types	Node types allowed in target nodes.