whoisjiji/fin-relevance

Financial-Relevance Classifier

The Financial-Relevance Classifier is a 0.6B-parameter binary text classifier for identifying finance-related questions and instructions. It is intended for high-throughput data mining and curation: given a user prompt, the model returns a continuous probability that the prompt belongs to the financial domain.

The model is introduced in Data-Centric Post-Training for Financial Reasoning: Mining, Distillation, and Verifiable Learning.

The positive class covers banking, investments, insurance, taxation, cryptocurrency, financial planning, corporate finance, economic theory, and personal finance. The model classifies domain relevance only; it does not evaluate answer correctness, self-containment, difficulty, or verifiability.

Model

The checkpoint adds a two-class sequence-classification head to Qwen/Qwen3-Embedding-0.6B. Inputs are user questions or instructions, truncated to at most 4,096 tokens. The semantic labels stored in config.json are:

Label	Meaning
NON_FINANCIAL	The prompt is outside the selected financial domains.
FINANCIAL	The prompt belongs to one of the selected financial domains.

Usage

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="whoisjiji/fin-relevance",
    tokenizer="whoisjiji/fin-relevance",
)

prompts = [
    "How does an increase in interest rates affect bond duration risk?",
    "Write a short poem about the ocean.",
]

scores = classifier(
    prompts,
    top_k=None,
    truncation=True,
    max_length=4096,
)

for prompt, result in zip(prompts, scores):
    probabilities = {item["label"]: item["score"] for item in result}
    print(prompt, probabilities["FINANCIAL"])

The returned value is a model score, not a calibrated universal decision boundary. Select a threshold using manually reviewed data from the intended deployment distribution. Lower thresholds favor recall for broad mining; higher thresholds favor precision for conservative filtering.

Limitations

• The labels were distilled from a generative model and may reproduce its domain biases.
• Adjacent areas such as economics, business, public policy, and quantitative reasoning can be ambiguous.
• The model was trained on English prompts and should not be assumed to generalize to other languages.
• A high financial-relevance score does not imply that a question is well-posed or that its answer is correct.
• This model is intended for dataset curation and research, not financial advice or automated high-stakes decisions.

Citation

@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}
}

@misc{hayrapetyan2026finrelevance,
  title  = {Financial-Relevance Classifier},
  author = {Hayrapetyan, Zhirayr},
  year   = {2026},
  url    = {https://huggingface.co/whoisjiji/fin-relevance}
}