IFRSLLM

IFRSLLM is a domain fine-tuned language model designed for experimentation with IFRS accounting, financial services, banking, and finance-related language tasks.

The model was fine-tuned using Supervised Fine-Tuning (SFT) with TRL and a domain-oriented instruction dataset containing approximately 20,000 examples.

The project explores how a relatively compact language model can be specialized for financial and accounting workflows and subsequently combined with Retrieval-Augmented Generation (RAG) to provide grounded access to authoritative and version-aware IFRS knowledge.

Important: IFRSLLM is an experimental AI model. It should not be treated as a replacement for professional accounting judgment, auditors, legal advisers, or authoritative IFRS materials.

Model

🤗 Hugging Face:

https://huggingface.co/eliashasnat/ifrsllm

Author: Elias Hasnat

Domain: IFRS, Accounting, Banking, Finance, Financial Services

Training method: Supervised Fine-Tuning (SFT)

Training framework: TRL + Transformers

Motivation

Financial and accounting language contains specialized terminology, structured reporting requirements, accounting standards, numerical information, and domain-specific reasoning patterns.

General-purpose language models may understand many of these concepts, but domain adaptation can improve the way a model interprets and responds to specialized financial instructions.

IFRSLLM explores a hybrid approach:

Domain Fine-Tuning + Authoritative RAG + Reranking + Version Control + Citations + Guardrails

The design principle behind the project is:

Fine-tuning teaches the model HOW to answer.

RAG provides WHAT is authoritative and current.

This separation is particularly important for IFRS because accounting standards, amendments, interpretations, effective dates, and jurisdiction-specific requirements may change over time.

Training Dataset

IFRSLLM was fine-tuned on approximately 20,000 instruction-response examples using an Alpaca-style structure:

{
  "instruction": "...",
  "input": "...",
  "output": "..."
}

The synthetic/domain-oriented dataset includes examples covering areas such as:

IFRS concept explanation
Financial statement summarization
IFRS 9 and Expected Credit Loss concepts
Accounting terminology
Banking customer service
Financial data interpretation
Financial risk indicators
PII redaction
Compliance-oriented tasks
Structured financial responses
Customer communication
Banking and finance question answering

The fine-tuning dataset is intended primarily to teach domain response patterns and instruction-following behavior.

It should not be considered an authoritative IFRS knowledge base.

Intended RAG Architecture

IFRSLLM is intended to work as the generation layer of a larger IFRS-focused RAG architecture.

User Question
      |
      v
Query Classification
      |
      v
Query Rewrite
      |
      v
+-----------------------------+
| Hybrid Retrieval            |
|                             |
| BM25 Search + Vector Search |
+-----------------------------+
      |
      v
Hybrid Fusion / RRF
      |
      v
Cross-Encoder Reranker
      |
      v
IFRS Context Builder
      |
      v
IFRSLLM
      |
      v
Grounding & Citation Validation
      |
      v
Structured IFRS Response

The external knowledge layer can contain appropriately licensed and authorized materials organized by:

IFRS standard
IAS standard
IFRIC interpretation
Topic
Paragraph
Effective date
Reporting period
Amendment version
Jurisdiction
Document type
Source

This architecture allows the model weights and accounting knowledge to evolve independently.

Example RAG Metadata

A retrieved IFRS knowledge chunk could use metadata similar to:

{
  "standard": "IFRS 9",
  "topic": "Expected Credit Loss",
  "section": "Impairment",
  "paragraph": "reference",
  "document_type": "standard",
  "effective_from": "YYYY-MM-DD",
  "jurisdiction": "IFRS",
  "language": "en",
  "source_version": "applicable-version",
  "source_url": "authorized-source"
}

This enables retrieval to consider both semantic relevance and accounting context.

Hybrid Retrieval

The planned retrieval architecture combines:

BM25 keyword retrieval

Useful for exact accounting terminology, standard names, and paragraph references.

Vector retrieval

Useful for semantic questions where users describe accounting situations without using exact IFRS terminology.

The results can then be combined using techniques such as Reciprocal Rank Fusion (RRF).

A cross-encoder reranker can further identify the passages most relevant to the accounting question before they are provided to IFRSLLM.

Version-Aware IFRS RAG

An important research direction for IFRSLLM is temporal accounting intelligence.

Instead of storing only the latest accounting information, the RAG layer can maintain multiple versions of standards and amendments.

Conceptually:

IFRS Knowledge Base

IFRS Standard
   |
   +-- Historical Version
   |
   +-- Current Version
   |
   +-- Amendments
   |
   +-- Future Effective Version

This could allow queries such as:

What accounting guidance was applicable to this transaction for the relevant reporting period?

The retrieval system can use the reporting date as part of the metadata filtering process before retrieving evidence.

Jurisdiction-Aware RAG

A future extension of IFRSLLM is jurisdiction-aware retrieval.

Conceptually:

                 Accounting Question
                         |
             +-----------+-----------+
             |                       |
             v                       v
       IFRS Knowledge          Jurisdiction
            Layer              Knowledge Layer
             |                       |
             +-----------+-----------+
                         |
                         v
                    IFRSLLM

This could help distinguish between:

IFRS requirements
Local accounting requirements
Regulatory requirements
Jurisdiction-specific adoption or implementation

These sources should remain clearly identified rather than being blended into an unsupported accounting conclusion.

Quick Start

from transformers import pipeline

model_id = "eliashasnat/ifrsllm"

generator = pipeline(
    "text-generation",
    model=model_id,
    device_map="auto"
)

question = """
Explain Expected Credit Loss under IFRS 9
in simple language.
"""

messages = [
    {
        "role": "user",
        "content": question
    }
]

output = generator(
    messages,
    max_new_tokens=256,
    return_full_text=False
)

print(output[0]["generated_text"])

Depending on the base model/chat template used during training, inference formatting may need to be adjusted accordingly.

Example Use Cases

Potential research and prototype use cases include:

IFRS Education

Explain accounting concepts in simpler language.

Financial Statement Analysis

Summarize structured financial information.

Accounting Knowledge Assistant

Combine IFRSLLM with RAG over authorized accounting materials.

Financial Services Assistant

Support domain-specific banking and finance questions.

IFRS Research Assistant

Retrieve relevant standards and present grounded summaries with references.

Multilingual Accounting Intelligence

Future RAG architectures can retrieve English authoritative sources while supporting questions and explanations in additional languages.

Recommended Production Architecture

For higher-stakes accounting use cases, IFRSLLM should not operate as a standalone source of accounting truth.

A stronger architecture is:

IFRSLLM
   +
Authoritative RAG
   +
Hybrid Retrieval
   +
Cross-Encoder Reranking
   +
Version Control
   +
Source Citations
   +
Grounding Validation
   +
Human Review

Limitations

IFRSLLM has several important limitations.

The model may:

Generate incorrect accounting information
Hallucinate IFRS requirements
Produce inaccurate paragraph references
Apply outdated accounting knowledge
Misinterpret complex transactions
Confuse similar accounting concepts
Produce mathematically incorrect results
Miss jurisdiction-specific requirements
Generate plausible but unsupported explanations

Fine-tuning does not guarantee factual correctness.

For this reason, authoritative accounting knowledge should preferably be supplied through a controlled RAG pipeline.

Responsible Use

IFRSLLM is intended for:

Research
Education
Experimentation
AI architecture development
RAG development
Financial NLP research
Prototype accounting assistants

It is not intended to provide independent professional accounting, audit, investment, tax, regulatory, or legal advice.

Important accounting conclusions should be verified against applicable authoritative standards and reviewed by qualified professionals.

IFRS Content and Licensing

This model is intended to learn domain-oriented response behavior rather than redistribute authoritative IFRS Standards.

Users building an IFRS RAG knowledge base are responsible for ensuring that source documents are obtained, stored, processed, and used under appropriate permissions and licensing arrangements.

IFRS and related marks and materials belong to their respective rights holders.

Evaluation

A future IFRSLLM evaluation suite is planned around metrics such as:

IFRS question-answer accuracy
Standard identification accuracy
Retrieval Recall@K
Citation accuracy
Groundedness
Hallucination rate
Numerical accuracy
Version accuracy
Jurisdiction accuracy
RAG faithfulness

The goal is to evaluate retrieval and generation separately so that errors can be traced to the retriever, reranker, context construction, or generation layer.

Future Roadmap

Planned research directions include:

Authoritative IFRS RAG
Hybrid BM25 + vector retrieval
Cross-encoder reranking
Citation verification
Version-aware retrieval
Jurisdiction-aware accounting intelligence
Multilingual IFRS queries
RAG evaluation framework
Agentic accounting workflows
Hallucination detection
Responsible AI guardrails
Human-in-the-loop accounting review

Training Procedure

The model was trained using Supervised Fine-Tuning (SFT) with TRL.

Framework Versions

TRL: 1.14.0
Transformers: 5.5.4
PyTorch: 2.14.0
Datasets: 5.0.1
Tokenizers: 0.22.2

Author

Elias Hasnat

AI | Enterprise Architecture | Financial Systems | LLM | RAG

Hugging Face:

https://huggingface.co/eliashasnat/ifrsllm

Citation

If you use IFRSLLM in research or experimentation, please reference the model repository:

@misc{hasnat2026ifrsllm,
  author       = {Elias Hasnat},
  title        = {IFRSLLM: Domain Fine-Tuned Language Model for IFRS and Financial Accounting},
  year         = {2026},
  howpublished = {Hugging Face},
  url          = {https://huggingface.co/eliashasnat/ifrsllm}
}

TRL

@software{vonwerra2020trl,
  title   = {{TRL: Transformers Reinforcement Learning}},
  author  = {von Werra, Leandro and others},
  license = {Apache-2.0},
  url     = {https://github.com/huggingface/trl},
  year    = {2020}
}

Disclaimer

IFRSLLM is an experimental research project.

Outputs should always be validated against applicable authoritative accounting standards and professional guidance before being used for accounting, audit, compliance, regulatory, investment, tax, or business decisions.

Downloads last month: -; Downloads are not tracked for this model. How to track