Fin-CFA-GraphGen: 785K Knowledge-Guided Financial QA Examples

Fin-CFA-GraphGen is a large-scale English dataset for financial instruction tuning, financial question answering, and domain-specific language-model post-training. It contains 785,149 synthetic question–answer examples generated from CFA curriculum and exam-preparation books with the GraphGen knowledge-driven data-generation method.

The source material covers core financial topics including economics, financial analysis, equity, fixed income, derivatives, alternative investments, portfolio management, and professional ethics. This is an intentionally broad candidate pool rather than a prescribed training split: every example includes continuous scores for financial relevance, self-containment, and answer verifiability so users can define selection thresholds for their own models and training stages.

Dataset construction

Source documents were divided into 1,024-token chunks with an overlap of 100 tokens. Qwen2.5-72B-Instruct was used in both stages of the GraphGen pipeline:

1. extracting entities, relations, and their descriptions into a fine-grained financial knowledge graph;
2. selecting connected graph neighborhoods and synthesizing questions and answers from the represented knowledge.

Generation used temperature 0, top-p 0.95, and a maximum of 8,192 output tokens. The generated prompts then underwent prompt-level semantic deduplication using Qwen3-Embedding-8B in a NeMo Curator workflow. The resulting broad pool contains 785,149 unique prompts.

No fixed classifier threshold is imposed in this release. The score columns are included to support application-specific filtering, ranking, curriculum construction, and selection of verifier-compatible tasks.

Generation methods

The method field records how the source graph was used to construct each example:

Method	Examples	Meaning
aggregated	422,487	Combines several related facts from a compact graph neighborhood into one question and answer.
multi_hop	337,875	Requires connecting information across multiple relations in a selected subgraph.
cot	24,787	Generated from a graph-guided reasoning path designed before producing the answer.

The method label describes the synthesis strategy. It does not guarantee that a separate chain-of-thought trace is present; this release stores the generated answer in a single assistant field.

Scoring models

The three score columns were produced by the following released classifiers:

Dataset field	Model	Interpretation
is_fin_score	Financial-Relevance Classifier	Probability that the prompt belongs to the selected financial domains.
is_selfcont_score	Lack-of-Context Detector	Self-containment score, computed as 1 - p(MISSING_CONTEXT).
is_verif_score	Verifiability Classifier	Probability that the answer has a compact, objectively checkable outcome.

The scores are already included in this dataset. The linked checkpoints can be used to score new data or reproduce the same selection signals. Each model card documents its label semantics, validation metrics, usage, and limitations.

Dataset structure

The repository contains fin-cfa-graphgen.jsonl (1,149,476,299 bytes). All 785,149 records are valid JSON and contain the following six non-empty fields:

Field	Type	Description
user	string	Synthetic financial question or instruction.
assistant	string	Synthetic answer generated from the selected graph context.
method	string	GraphGen synthesis strategy: aggregated, multi_hop, or cot.
is_fin_score	float	Predicted financial relevance; larger values indicate stronger finance-domain relevance.
is_selfcont_score	float	Predicted self-containment; larger values favor questions answerable from the information provided. It equals one minus the predicted missing-context probability.
is_verif_score	float	Predicted answer verifiability; larger values favor compact, objectively checkable outcomes. This is not an answer-correctness score.

Loading

from datasets import load_dataset

dataset = load_dataset(
    "whoisjiji/fin-cfa-graphgen",
    split="train",
    streaming=True,
)
example = next(iter(dataset))

The train designation is a packaging convention; no official train/validation/test split is provided. Users should create their own held-out split after applying any task-specific selection and deduplication rules.

Selecting examples

The three classifier outputs are continuous selection signals, not gold labels. There is no universally correct threshold:

• is_fin_score supports stricter or broader definitions of financial relevance;
• is_selfcont_score helps reject questions that may depend on missing source context;
• is_verif_score can help identify tasks whose answers are suitable for compact rule-based or parser-based rewards.

Thresholds should be calibrated through manual inspection and validation against the intended model, verifier, and downstream task distribution. The fields can also be used for ranking or curriculum construction instead of hard filtering.

Intended uses

• Financial instruction tuning and continued post-training
• Financial question answering and concept-coverage studies
• Construction of high-precision SFT subsets
• Mining self-contained or verifier-compatible tasks for RL and GRPO
• Research on knowledge-graph-guided synthetic data generation
• Studying stage-specific data selection with learned classifiers

Quality, risks, and limitations

The dataset was structurally validated for JSON parsing, field presence, and method values. This does not guarantee factual correctness. Questions and answers are synthetic and may contain incorrect calculations, unsupported claims, ambiguous wording, or information that has become outdated. Classifier scores are model predictions and should not be interpreted as human judgments or correctness guarantees.

The release contains generated questions and answers, not per-example source passages or page-level citations, so many answers cannot be traced directly to a particular passage. No benchmark-decontamination guarantee is provided. Users should check overlap with their evaluation sets, manually review selected samples, and independently verify answers before high-stakes use.

This is not an official CFA Institute dataset and is not affiliated with or endorsed by CFA Institute. The dataset is intended for research and model development and does not constitute financial advice.

Citation

If you use this dataset, please cite both this release and the GraphGen method:

@dataset{hayrapetyan2026fincfagraphgen,
  title  = {Fin-CFA-GraphGen: 785K Knowledge-Guided Financial QA Examples},
  author = {Hayrapetyan, Zhirayr},
  year   = {2026},
  url    = {https://huggingface.co/datasets/whoisjiji/fin-cfa-graphgen}
}

@article{chen2025graphgen,
  title   = {GraphGen: Enhancing Supervised Fine-Tuning for LLMs with Knowledge-Driven Synthetic Data Generation},
  author  = {Chen, Zihong and Jiang, Wanli and Li, Jinzhe and Yuan, Zhonghang and Kong, Huanjun and Ouyang, Wanli and Dong, Nanqing},
  journal = {arXiv preprint arXiv:2505.20416},
  year    = {2025}
}