VANGUARD-MODEL-V1

VANGUARD-MODEL-V1

Evidence-Grounded Financial Intelligence for Enterprise Applications

VANGUARD-MODEL-V1 is a financial domain-adapted language model based on Qwen3-4B-Instruct-2507, fine-tuned using Quantized Low-Rank Adaptation (QLoRA) for financial question answering and document reasoning.

Designed for integration into enterprise financial intelligence systems, the model aims to help applications answer questions about financial statements, annual reports, SEC filings, corporate disclosures, and financial tables using supplied evidence.

The model is intended to operate within a retrieval-augmented generation (RAG) pipeline, with source retrieval, citation validation, and deterministic numerical computation handled by the surrounding application where required.

Core objective: Make financial document question answering more structured, traceable, and suitable for verification in enterprise workflows.

«Important: VANGUARD-MODEL-V1 is a language-model component, not a complete financial intelligence platform. Its outputs can contain errors. Financial calculations, citations, and consequential conclusions should be independently verified.»

Model Details

Model Description

• Model name: VANGUARD-MODEL-V1
• Developer: "Md Asif" (https://huggingface.co/Md-Asif)
• Project: VANGUARD.AI
• Model type: Causal language model with parameter-efficient fine-tuning
• Base model: "Qwen/Qwen3-4B-Instruct-2507" (https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507)
• Fine-tuning method: QLoRA with supervised fine-tuning (SFT)
• Primary language: English
• Primary domain: Financial document question answering and reasoning
• Model parameters: Approximately 4.06 billion in the base model
• Model format: PEFT adapter, if the repository contains adapter weights
• License: MIT, as declared for this repository
• Intended users: Financial enterprises, developers, AI engineers, and financial NLP researchers
• Project repository: "github.com/MdAsif-Ai/vanguard.ai" (https://github.com/MdAsif-Ai/vanguard.ai)
• Research paper: Not yet published

Model Sources

• Hugging Face developer profile: https://huggingface.co/Md-Asif
• Base model: https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507
• GitHub repository: https://github.com/MdAsif-Ai/vanguard.ai
• Demo: No standalone public demo specified.

Intended Uses

Direct Use

VANGUARD-MODEL-V1 is intended for financial question answering when relevant documents, passages, or structured financial data are provided as input.

Potential tasks include:

• Financial statement analysis: Answer questions about revenue, operating income, net income, expenses, debt, and margins.
• Corporate filing analysis: Extract and interpret information from annual reports, SEC 10-K and 10-Q filings, and corporate disclosures.
• Numerical reasoning: Support questions requiring calculations over financial tables and reported metrics.
• Evidence-grounded responses: Generate answers referencing numbered evidence passages supplied in the prompt.
• Financial document reasoning: Answer questions that require combining relevant information from multiple supplied passages.
• Answerability assessment: Identify questions that cannot be answered adequately from the supplied evidence, subject to validation.

Downstream Use

The model can be integrated into enterprise applications and document intelligence systems, including:

• Financial document intelligence platforms.
• Enterprise RAG applications.
• Financial reporting and analysis workflows.
• Internal research and document search systems.
• Financial question-answering evaluation pipelines.
• Audit-support applications requiring traceable source references.
• Self-hosted financial AI deployments.

For enterprise deployments, the model should be combined with document retrieval, access control, evidence validation, deterministic calculation tools, logging, and human review where appropriate.

Out-of-Scope Use

VANGUARD-MODEL-V1 is not intended to be used as:

• A replacement for qualified financial analysts, accountants, auditors, or legal professionals.
• A source of real-time market data without an external data connection.
• A stock-price prediction or investment recommendation model.
• An autonomous system for approving loans, executing trades, or making other consequential financial decisions without appropriate safeguards.
• A guaranteed source of accurate financial calculations.
• A substitute for source documents, accounting controls, or regulatory compliance procedures.
• A general-purpose assistant whose reliability has been established across unrelated domains.

Bias, Risks, and Limitations

Technical Limitations

1. Hallucinations

The model may generate unsupported statements, incorrect financial figures, or plausible-sounding explanations. Providing evidence reduces reliance on unsupported knowledge but does not eliminate hallucinations.

1. Numerical accuracy

Language models can make arithmetic errors, misinterpret percentages, confuse units, or compare financial values from incompatible reporting periods. Use deterministic calculation tools for material financial computations.

1. Citation reliability

A generated citation does not guarantee that the cited passage supports the associated claim. Applications should validate citations against the original retrieved passages.

1. Retrieval dependence

The quality of a RAG-based answer depends on document extraction, chunking, retrieval, ranking, and the relevance and completeness of the supplied evidence.

1. Domain generalization

Evaluation on financial benchmarks does not establish equivalent performance across all industries, accounting standards, jurisdictions, document formats, or languages.

1. Context length

The reported training sequence length is 1,024 tokens. Longer inference contexts should be evaluated separately, and effective context length depends on the base model, serving configuration, and application.

1. Model scale

The approximately 4-billion-parameter base model is designed for a more resource-efficient deployment profile than larger models, but it may struggle with complex multi-document reasoning and subtle financial interpretations.

1. Temporal limitations

The model does not inherently provide live financial information. Current figures and events must be retrieved from reliable, authorized sources.

Security and Enterprise Risks

• Retrieved documents may contain malicious or misleading instructions.
• Incorrect source permissions can expose confidential financial information.
• Model outputs may omit important qualifications or reporting-period distinctions.
• Unsupported citations may create a false impression of auditability.
• Incorrect answers may lead to financial or operational losses if used without verification.

Recommendations

For enterprise deployment:

1. Supply relevant evidence from authorized source documents.
2. Enforce document-level permissions before retrieval and generation.
3. Use Python or another deterministic calculation engine for critical numerical operations.
4. Validate every citation against its source passage.
5. Require the model to acknowledge insufficient evidence.
6. Test prompt-injection resistance and document isolation.
7. Maintain appropriate audit logs and versioned evaluation datasets.
8. Benchmark on representative enterprise documents before deployment.
9. Require human approval for consequential financial decisions.
10. Reevaluate performance whenever the model, retrieval pipeline, prompts, or source data change.

How to Get Started with the Model

1. Install dependencies

pip install -U torch transformers peft accelerate

1. Load the base model and fine-tuned adapter

The following example assumes the Hugging Face repository contains a PEFT adapter and that its repository ID is "Md-Asif/VANGUARD-MODEL-V1".

import torch from transformers import ( AutoModelForCausalLM, AutoTokenizer, ) from peft import PeftModel

BASE_MODEL = "Qwen/Qwen3-4B-Instruct-2507" ADAPTER_MODEL = "Md-Asif/VANGUARD-MODEL-V1"

tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)

base_model = AutoModelForCausalLM.from_pretrained( BASE_MODEL, torch_dtype="auto", device_map="auto", )

model = PeftModel.from_pretrained( base_model, ADAPTER_MODEL, )

model.eval()

messages = [ { "role": "system", "content": ( "You are a financial document assistant. " "Answer using the supplied evidence. " "Cite relevant passages using their numbered references. " "If the evidence is insufficient, state that clearly. " "Do not invent financial figures." ), }, { "role": "user", "content": ( "Evidence:\n" "[1] Revenue for the reporting period was $400 million.\n" "[2] Operating income for the same period was $80 million.\n\n" "Question: What was the operating margin?" ), }, ]

prompt = tokenizer.apply_chat_template( messages, tokenize=False, add_generation_prompt=True, )

inputs = tokenizer( prompt, return_tensors="pt", ).to(model.device)

with torch.inference_mode(): outputs = model.generate( **inputs, max_new_tokens=160, do_sample=False, )

new_tokens = outputs[0][inputs["input_ids"].shape[1]:]

answer = tokenizer.decode( new_tokens, skip_special_tokens=True, )

print(answer)

The example uses illustrative figures. For production financial applications, calculate the operating margin with a deterministic tool and verify that the source passages support the answer.

Loading note: If you publish a merged full model rather than a PEFT adapter, follow the loading instructions appropriate to that artifact. The base model and adapter must be compatible.

1. Serve using vLLM

For compatible versions of vLLM and PEFT, you can try serving the adapter with:

vllm serve Md-Asif/VANGUARD-MODEL-V1 \ --gpu-memory-utilization 0.80 \ --max-model-len 4096

Adapter-serving support and configuration depend on the installed vLLM version and the repository artifacts. If this command is not compatible with your published model, merge the adapter with its base model and serve the merged directory instead.

1. Recommended enterprise architecture

A production financial intelligence system should generally separate language generation from retrieval, calculation, and verification.

Suggested workflow:

1. Ingest and parse authorized financial documents.
2. Retrieve relevant evidence using hybrid search and metadata filters.
3. Provide the retrieved evidence and question to VANGUARD-MODEL-V1.
4. Execute critical calculations using deterministic code.
5. Validate generated citations and numerical claims.
6. Return the answer with source references.
7. Record appropriate audit metadata and evaluation outcomes.

The model alone does not implement this complete workflow.

Training Details

Training Data

The reported training-data sources and example counts are listed below. These counts describe the reported dataset composition and should be reconciled against the final, deduplicated training manifest before being treated as the number of unique examples.

Source| Reported examples FinQA training split| 4,000 ConvFinQA training split| 3,000 TAT-QA training split| 2,000 Synthetic QA from SEC 10-K filings| 3,000 Chain-of-thought augmented examples| 2,000 Calculation problems| 2,000 Adversarial and distractor examples| 2,000

Important dataset accounting note: These category counts sum to 18,000. The original training description states 12,000 examples. The categories may overlap, or one of the reported totals may be incorrect. The final unique example count has not been established by these figures alone.

Data Preprocessing and Validation

The reported training approach includes:

• Chat-style message formatting.
• Numbered evidence passages.
• Financial question-and-answer examples.
• Numerical reasoning and calculation problems.
• Distractor-evidence examples.
• Examples involving insufficient evidence and adversarial inputs.
• Schema validation, evidence-support checks, numerical verification, and deduplication.

The following claims should be supported by reproducible data audits before publication: the percentage of Python-verified calculations, the proportion of unanswerable and adversarial examples, and the absence of benchmark leakage.

Training datasets may have their own licenses and redistribution requirements. Verify that the use of each dataset and any derived examples complies with its applicable terms.

Training Procedure

The model was reportedly fine-tuned using QLoRA with supervised fine-tuning.

Hyperparameter| Reported configuration Training regime| BF16 compute with QLoRA Quantization| 4-bit NF4 Double quantization| Enabled LoRA rank| 16 LoRA alpha| 32 LoRA dropout| 0.05 Target modules| q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj Learning rate| 2e-4 Scheduler| Cosine Warmup ratio| 0.03 Per-device batch size| 1 Gradient accumulation steps| 16 Effective batch size| Depends on GPU count and distributed configuration Epochs| 3 Maximum sequence length| 1,024 tokens Optimizer| paged_adamw_8bit Training framework| Transformers, TRL, PEFT, bitsandbytes

Training Hardware and Performance

The following are reported project measurements and should be checked against the actual training logs and artifacts.

Measurement| Reported value GPU| NVIDIA GeForce RTX 5070 GPU memory| Approximately 12 GB Peak training VRAM| 6.34 GB Training duration| Approximately 3.5 hours Training loss| 2.688 Validation loss| 2.305 Adapter size| Approximately 66.1 MB Merged model size| Approximately 7.50 GB

Actual artifact sizes may vary with serialization format, quantization, and precision.

Evaluation

Testing Data

The reported evaluation uses three financial QA benchmarks:

Benchmark| Reported test size| Task FinanceBench| 150| Financial question answering over corporate filings FinQA test| 1,147| Numerical reasoning over financial tables TAT-QA test| 1,663| Question answering over tables and text Total| 2,960| Combined evaluation questions

For reproducibility, specify the dataset versions, exact splits, prompt templates, retrieval configuration, answer-normalization procedure, and evaluation scripts.

Evaluation Factors

The reported evaluation dimensions include:

• Question type: extraction, calculation, multi-hop reasoning, unanswerable questions, and adversarial examples.
• Difficulty: easy, medium, and hard.
• Company and industry coverage.

The number of companies, sectors, and examples in each evaluation subgroup should be documented before publishing subgroup-specific conclusions.

Evaluation Metrics

• Numerical accuracy: Correctness of numerical answers under a defined tolerance.
• Exact match: Exact agreement with the normalized reference answer.
• Token-level F1: Overlap between predicted and reference answer tokens.
• Citation accuracy: Whether generated citations refer to relevant supporting evidence.
• Evidence grounding: Whether answer claims are supported by the supplied documents.
• Hallucination rate: Frequency of unsupported claims under the chosen evaluation procedure.
• Abstention accuracy: Ability to refrain from answering when evidence is insufficient.
• Inference latency: Time required to generate an answer under a specified serving configuration.
• Completion length: Number of generated tokens per answer.

Reported Benchmark Results

The following table preserves the reported results. Treat them as project-reported results until the underlying evaluation artifacts are available for independent verification.

Experiment| Configuration| FinanceBench| FinQA test| TAT-QA test| Reported overall E0| Base model, no RAG| 54.7%| 5.8%| 21.6%| 17.2% E1| Base model + RAG| 87.3%| 44.4%| 76.1%| 64.4% E5| Fine-tuned model + RAG| 83.3%| 43.6%| 76.1%| 63.9%

Interpretation: In the reported results, the fine-tuned model with RAG scored 0.5 percentage points lower on the overall metric than the base model with RAG. This difference alone does not establish statistical equivalence or demonstrate that fine-tuning improves accuracy.

The overall metric requires an explicit definition because it does not appear to be a simple arithmetic average of the three displayed benchmark scores.

Additional Reported Measurements

Measurement| Fine-tuned configuration| Base configuration Mean inference latency| 3,873 ms| 4,162 ms Mean completion length| 230.9 tokens| 241.7 tokens

Additional reported results:

• Citation accuracy: 81.6%.
• Abstention failures: zero in the reported evaluation.

These measurements require documented scoring procedures and test conditions. In particular, the abstention result should include the number and composition of unanswerable test cases.

Evaluation Limitations

Benchmark performance does not guarantee performance on an enterprise's internal financial documents. Real-world deployment may involve different document layouts, OCR errors, accounting conventions, retrieval quality, reporting periods, and access-control requirements.

Users should perform independent evaluations on representative data before making deployment decisions.

Environmental Impact

Carbon emissions have not been independently measured for this model. An estimate requires the energy consumed during training and evaluation, the hardware power profile, and the applicable electricity-grid emissions factor.

Reported compute information:

• Hardware: NVIDIA GeForce RTX 5070, approximately 12 GB VRAM.
• Training duration: Approximately 3.5 hours.
• Evaluation duration: Approximately 0.5 hours, as reported.
• Cloud provider: None reported; local compute.
• Compute region: India.

An estimate of approximately 0.22 kg CO₂-equivalent was mentioned in the original project notes, but its calculation has not been independently established. It should not be presented as a verified emissions measurement without supporting assumptions and methodology.

For estimation methodology, see the "Machine Learning Impact calculator" (https://mlco2.github.io/impact#compute) and "Lacoste et al. (2019)" (https://arxiv.org/abs/1910.09700).

Technical Specifications

Model Architecture and Objective

• Base architecture: Qwen3-4B-Instruct-2507.
• Architecture type: Dense causal transformer.
• Fine-tuning method: QLoRA.
• Training objective: Supervised fine-tuning.
• Reported training sequence length: 1,024 tokens.
• Intended inference context: Up to 4,096 tokens in the reported serving configuration, subject to base-model and serving compatibility.
• Primary task: Text generation for financial question answering.

Compute Infrastructure

Reported hardware

• GPU: NVIDIA GeForce RTX 5070.
• GPU memory: Approximately 12 GB.
• System RAM: 32 GB.
• CPU: 40 cores, as reported.

Reported software stack

• PyTorch.
• Transformers.
• TRL.
• PEFT.
• bitsandbytes.

Exact package versions should be recorded from the actual training environment to support reproducibility.

License

This repository declares the MIT License.

The MIT License applies to this repository's contents to the extent that the publisher has the rights to license them. It does not automatically change the license or terms governing the base model, third-party datasets, or other dependencies.

Review the base model's terms and the applicable dataset licenses before redistribution or commercial deployment.

Citation

No published paper has been provided for VANGUARD-MODEL-V1. Do not cite the model as a peer-reviewed publication until an actual paper has been published.

For now, reference the model and project directly:

APA-style project reference

Asif, M. (n.d.). VANGUARD-MODEL-V1: Financial domain-adapted language model. Hugging Face. https://huggingface.co/Md-Asif/VANGUARD-MODEL-V1

BibTeX

@misc{asif_vanguard_model_v1, author = {Md Asif}, title = {VANGUARD-MODEL-V1: Financial Domain-Adapted Language Model}, howpublished = {Hugging Face}, url = {https://huggingface.co/Md-Asif/VANGUARD-MODEL-V1}, note = {Model card and implementation details} }

Update the citation if the repository name changes or a paper is published.

Glossary

• QLoRA: Quantized Low-Rank Adaptation, a parameter-efficient fine-tuning approach.
• SFT: Supervised fine-tuning using example input-output pairs.
• RAG: Retrieval-augmented generation, which supplies relevant external evidence to a language model.
• PEFT: Parameter-Efficient Fine-Tuning.
• FinanceBench: A benchmark for evaluating financial question answering over company filings.
• FinQA: A benchmark for numerical reasoning over financial reports and tables.
• TAT-QA: A benchmark for question answering that combines tabular and textual evidence.
• Evidence grounding: The degree to which an answer is supported by its cited sources.
• Abstention: Declining to provide an answer when sufficient evidence is unavailable.

More Information

• Hugging Face profile: https://huggingface.co/Md-Asif
• GitHub project: https://github.com/MdAsif-Ai/vanguard.ai
• Base model: https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507

Model Card Authors

Md Asif — independent developer of VANGUARD.AI.

Model Card Contact

For issues, technical questions, and contributions, use the GitHub repository:

https://github.com/MdAsif-Ai/vanguard.ai