NurErtug/pit-finance-grpo-merged

PIT Finance GRPO + Verifier System

This repository contains a verifier-guarded financial question answering system built on top of the PIT-4B point-in-time financial language model family.

The project combines:

• A merged GRPO-aligned PIT generator
• Numeric verification
• NLI-based semantic grounding verification
• Abstention behavior for unsupported answers

The objective is to improve reliability and reduce hallucinations in financial QA over earnings-call transcripts and financial documents.

---

Model Overview

Base Model

• Diamegs/PIT-4B-FT-202212

Alignment

The model was aligned using:

• reasoning supervision
• GRPO-style preference optimization
• LoRA fine-tuning
• hallucination-aware financial QA data

The LoRA adapter was merged into the final model weights.

Final Merged Model

• NurErtug/pit-finance-grpo-merged

---

Intended Use

The system is designed for:

• grounded financial QA
• earnings-call reasoning
• hallucination reduction
• numeric consistency
• abstention on unsupported questions
• verifier-guided inference

---

Verifier Architecture

The verifier is implemented as an inference-time guardrail wrapper.

The verifier is NOT merged into model weights.

The pipeline operates as follows:

User Question
      ↓
Merged PIT Generator
      ↓
Numeric Verifier
      ↓
NLI Verifier
      ↓
Abstention Policy
      ↓
Final Answer

---

Verification Components

1. Numeric Verifier

Checks whether numeric claims generated by the model are supported by the evidence/context.

Examples:

• unsupported dollar amounts
• hallucinated percentages
• fabricated growth figures

---

2. NLI Verifier

Uses:

• MoritzLaurer/DeBERTa-v3-base-mnli-fever-anli

to verify semantic support between:

• evidence/context
• generated answer

Possible outcomes:

• entailment
• neutral
• contradiction

---

3. Abstention Policy

If the verifier determines:

• insufficient evidence
• unsupported claims
• semantic contradiction

the system returns:

NOT ENOUGH INFORMATION.

instead of hallucinating an answer.

---

Installation

Clone the repository:

git clone https://huggingface.co/NurErtug/pit-finance-grpo-merged
cd pit-finance-grpo-merged

Install dependencies:

pip install -r requirements.txt

---

Loading the Model

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_name = "NurErtug/pit-finance-grpo-merged"

tokenizer = AutoTokenizer.from_pretrained(
    model_name,
    trust_remote_code=True,
    use_fast=False,
)

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    trust_remote_code=True,
    torch_dtype=torch.float16,
    device_map="auto",
)

model.eval()

---

Using the Verifier Wrapper

from verifier_wrapper import VerifierWrapper

wrapper = VerifierWrapper()

context = """
Revenue increased 18% year-over-year.
Alternative minimum tax credits totaled $125 million.
"""

question = "What was the revenue growth?"

result = wrapper.answer(
    context=context,
    question=question,
)

print(result)

---

Example Output

{
    "initial_answer": "Revenue increased 18% year-over-year.",
    "final_answer": "Revenue increased 18% year-over-year.",
    "rejected": False,
    "reward": 1.8
}

Example rejected output:

{
    "initial_answer": "$250 million",
    "final_answer": "NOT ENOUGH INFORMATION.",
    "rejected": True,
    "reject_reason": "verifier_failed"
}

---

Repository Contents

README.md
verifier_wrapper.py
requirements.txt
config.json
generation_config.json
model.safetensors
tokenizer files

---

Research Motivation

Financial QA systems are especially vulnerable to:

• hallucinated financial figures
• unsupported claims
• temporal leakage
• fabricated reasoning

This project explores verifier-guided inference as a lightweight alternative to:

• expensive RLHF pipelines
• larger judge-model architectures
• fully integrated verifier training

The verifier wrapper provides an interpretable post-generation filtering mechanism for improving reliability in financial language models.

---

Limitations

The system may still:

• over-abstain
• reject partially correct answers
• fail under highly ambiguous contexts
• struggle with extremely long contexts

The verifier improves robustness but does not guarantee factual correctness.

---

Citation

@misc{pit_finance_verifier_2026,
  title={PIT Finance GRPO + Verifier System},
  author={Nur Ertug},
  year={2026},
  publisher={Hugging Face}
}

---

License

Apache-2.0