Model Card for fenra-V1

Model Overview

fenra-V1 is a domain-specialized language model focused on procurement fraud detection, analysis, and related investigative tasks. It is currently under active development and fine-tuning using parameter-efficient techniques.

The model is designed to assist with identifying suspicious procurement patterns, generating investigative insights, and supporting analysts working in fraud detection and compliance domains.

⚠️ Note: Training is ongoing. The model is not yet merged and may produce unstable or inconsistent outputs.

---

Model Details

Model Description

fenra-V1 is a fine-tuned variant of Phi-3-medium-4k-instruct, adapted using QLoRA (Quantized Low-Rank Adaptation) for efficient training and deployment. The model leverages domain-specific data related to procurement fraud scenarios to enhance its contextual understanding and response quality in this niche.

• Developed by: Fenra Project
• Model type: Causal Language Model (Instruction-tuned)
• Base model: unsloth/Phi-3-medium-4k-instruct
• Fine-tuning method: QLoRA (parameter-efficient fine-tuning)
• Training status: Ongoing (LoRA adapters not yet merged)
• Language(s): English
• License: MIT

---

Model Sources

• Base Model: https://huggingface.co/unsloth/Phi-3-medium-4k-instruct
• Training Dataset: https://huggingface.co/datasets/monadgeek/fenra/
• Repository: To be added
• Demo: To be added

---

Intended Uses

Direct Use

fenra-V1 can be used for:

• Procurement fraud analysis
• Risk flagging in procurement documents
• Generating investigative summaries
• Question answering within fraud/compliance contexts
• Pattern recognition in suspicious transactions or tenders

Downstream Use

• Integration into fraud detection platforms
• Compliance and audit tooling
• Risk scoring pipelines
• Internal investigation assistants

Out-of-Scope Use

This model is not suitable for:

• Legal decision-making without human oversight
• Financial or regulatory compliance automation without validation
• High-stakes decision systems
• General-purpose reasoning outside its domain
• Use cases requiring guaranteed factual accuracy

---

Bias, Risks, and Limitations

Limitations

• Training is ongoing; outputs may be inconsistent
• Domain bias toward procurement fraud scenarios
• May hallucinate or fabricate details
• Limited general-world knowledge beyond training scope
• Not evaluated for fairness across demographic groups

Risks

• Misinterpretation of model outputs as factual evidence
• Over-reliance in investigative workflows
• Potential false positives in fraud detection scenarios

Recommendations

• Always use human-in-the-loop verification
• Treat outputs as assistive, not authoritative
• Validate findings with external data sources
• Avoid use in automated enforcement or legal systems

---

Getting Started

Example usage (Transformers + PEFT):

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model = "unsloth/Phi-3-medium-4k-instruct"
lora_adapter = "your-fenra-lora-path"

tokenizer = AutoTokenizer.from_pretrained(base_model)
model = AutoModelForCausalLM.from_pretrained(base_model)

model = PeftModel.from_pretrained(model, lora_adapter)

prompt = "Analyze this procurement record for fraud indicators: ..."
inputs = tokenizer(prompt, return_tensors="pt")

outputs = model.generate(**inputs, max_new_tokens=200)
print(tokenizer.decode(outputs[0]))