whoisjiji/fin-glaive

Fin-Glaive: 645K Financial Instruction and Reasoning Examples

Fin-Glaive is a large-scale English dataset for financial instruction tuning, financial question answering, and reasoning-focused language-model post-training. It contains 645,232 question–reasoning–answer examples mined from Glaive Reasoning v1 20M.

The dataset and its role in the post-training pipeline are described in Data-Centric Post-Training for Financial Reasoning: Mining, Distillation, and Verifiable Learning.

The dataset is designed for supervised fine-tuning (SFT), chain-of-thought and reasoning distillation, financial-domain adaptation, and constructing reinforcement-learning or GRPO subsets. Every example includes three continuous selection scores for financial relevance, self-containment, and answer verifiability, allowing users to choose thresholds appropriate for their own training stage.

Dataset construction

Finance-related questions were mined from the broad Glaive Reasoning v1 corpus with a lightweight financial-relevance classifier. The candidate pool then underwent prompt-level semantic deduplication and response parsing, yielding the 645,232 examples in this release. Reasoning traces and final answers are stored separately.

Rather than imposing one universal quality threshold, this release exposes the auxiliary scores used by the curation pipeline. This makes it possible to build broad financial instruction mixtures, conservative SFT subsets, or narrower verifier-compatible RL datasets from the same pool.

Scoring models

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

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 one JSONL file, fin-glaive.jsonl (5,327,140,792 bytes). All 645,232 records contain the following six non-empty fields:

Loading

from datasets import load_dataset


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

Streaming is recommended when exploring the full 5.33 GB JSONL file on a memory-constrained machine. The train designation is a packaging convention; no official evaluation split is provided.

Building a stricter financial SFT subset

The released scores can be combined to create a conservative starting subset. Thresholds are application-dependent; the following example is illustrative rather than a prescribed benchmark split:

from datasets import load_dataset


dataset = load_dataset("whoisjiji/fin-glaive", split="train", streaming=True)
strict_finance = dataset.filter(
    lambda row: row["is_fin_score"] > 0.99
    and row["is_selfcont_score"] > 0.90
)

For compact rule-based RL rewards, is_verif_score can be used as an additional ranking or filtering signal. Users should select and validate the threshold against their own verifier and target task distribution.

Intended uses

• —Financial instruction tuning and continued post-training
• —Financial question answering and long-form reasoning
• —Chain-of-thought and reasoning-trace research
• —Domain-specific SFT dataset construction
• —Mining verifiable finance tasks for RL, RLHF, or GRPO
• —Studying learned data filters and stage-specific data selection

Quality, risks, and limitations

All records were checked for JSON parsing and the presence of the six fields. This structural validation does not guarantee factual correctness. The source responses are synthetically generated and may contain incorrect calculations, unsupported claims, or flawed reasoning. The three scores are classifier predictions, not human annotations or correctness guarantees.

The release does not establish benchmark decontamination or a dedicated privacy audit. Users should check overlap with their evaluation sets, review samples selected by their thresholds, and independently validate answers before high-stakes use. The dataset is intended for research and model development and does not constitute financial advice.

Source, attribution, and license

The source dataset is Glaive AI's [Reasoning-v1-20m](https://huggingface.co/datasets/glaiveai/reasoning-v1-20m), released under Apache-2.0. This derivative release preserves the Apache-2.0 designation and adds finance-oriented mining, semantic deduplication, a separated question/reasoning/answer schema, and auxiliary classifier scores.

When using this dataset, please credit both the original Glaive dataset and this curated release:

@article{hayrapetyan2026datacentric,
  title   = {Data-Centric Post-Training for Financial Reasoning: Mining, Distillation, and Verifiable Learning},
  author  = {Hayrapetyan, Zhirayr and Kalmykov, Andrei and Kokosinskii, Denis and Stanishevskii, Dmitry and Zmitrovich, Dmitry},
  journal = {arXiv preprint arXiv:2609.10113},
  year    = {2026}
}

@dataset{hayrapetyan2026finglaive,
  title  = {Fin-Glaive: 645K Financial Instruction and Reasoning Examples},
  author = {Hayrapetyan, Zhirayr},
  year   = {2026},
  url    = {https://huggingface.co/datasets/whoisjiji/fin-glaive}
}