theeseus-ai/CriticalThinkerRisk-Base-8B

RiskClassifier: Fine-Tuned LLaMA 3.1 8B Model

Model Summary

RiskClassifier is a fine-tuned version of the meta-llama/Llama-3.1-8B-Instruct model, designed to evaluate risk levels across diverse scenarios using structured critical thinking. It is fine-tuned on the theeseus-ai/RiskClassifier dataset, which focuses on assessing and labeling risk scores while maintaining detailed reasoning explanations. This model is optimized for tasks requiring risk classification, fraud detection, and analytical reasoning.

Model Details

• Base Model: meta-llama/Llama-3.1-8B-Instruct
• Fine-tuned Dataset: theeseus-ai/RiskClassifier
• Model Size: 8 Billion Parameters
• Language: English
• License: Apache 2.0
• Use Case: Risk assessment, fraud detection, critical thinking tasks

Dataset Information

The RiskClassifier dataset provides structured scenarios with:

• Context: A description of the event requiring analysis.
• Query: A critical-thinking question tied to the scenario.
• Answers: Four risk level options ("Low risk," "Moderate risk," "High risk," "Very high risk").
• Risk Score: A numeric value (0–100) representing the raw risk assessment.
• Conversations: Reformatted data in ShareGPT-style conversation format to train the model for reasoning and structured responses.

Example Reformatted Output:

{
  "context": "A customer used a credit card in a high-fraud region for a large purchase.",
  "query": "What is the risk level of this transaction?",
  "answers": ["Low risk", "Moderate risk", "High risk", "Very high risk"],
  "risk_score": 85,
  "conversations": [
    {"role": "system", "content": "You are a helpful AI that assesses risk levels and provides explanations."},
    {"role": "user", "content": "Context: A customer used a credit card in a high-fraud region for a large purchase.\nQuestion: What is the risk level of this transaction?\nAnswers: [Low risk, Moderate risk, High risk, Very high risk]"},
    {"role": "assistant", "content": "Risk Level: Very high risk (Score: 85)"}
  ]
}

Intended Use

Applications

• Fraud Detection: Evaluating suspicious transactions and identifying high-risk activities.
• Risk Analysis: Assessing scenarios with probabilistic evaluations for financial and operational decisions.
• Critical Thinking Tasks: Enhancing AI's ability to reason about uncertainty and complex situations.
• Educational Tools: Training AI systems to provide explanations for risk assessments.

Limitations

• Context Dependency: Accuracy may degrade with ambiguous or incomplete context.
• Bias Risk: Outputs may inherit biases present in training data; manual review is advised for high-impact decisions.
• Numeric Risk Scores: The numerical scores may require post-processing to fit domain-specific thresholds.

How to Use

Example Code:

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_name = "theeseus-ai/RiskClassifier"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

inputs = tokenizer("Context: A large transaction flagged for manual review.\nQuestion: What is the risk level?", return_tensors="pt")
outputs = model.generate(**inputs, max_length=100)
print(tokenizer.decode(outputs[0]))

Evaluation Metrics

• Accuracy: Verified predictions against labeled risk levels.
• Reasoning Completeness: Evaluated explanations for clarity and alignment with context.
• Risk Score Consistency: Checked correlation between numeric risk scores and label predictions.

Training Configuration

• Optimizer: AdamW
• Batch Size: 32
• Learning Rate: 2e-5
• Epochs: 3
• Hardware: NVIDIA A100 GPUs
• Precision: bf16 mixed precision

Environmental Impact

• Hardware: NVIDIA A100 GPUs
• Training Hours: ~2 hours
• Carbon Emissions: Estimated using ML CO2 Calculator

Citation

@misc{RiskClassifier2024,
  title={RiskClassifier: Fine-Tuned LLaMA 3.1 8B Model for Risk Assessment},
  author={Theeseus AI},
  year={2024},
  howpublished={\url{https://huggingface.co/theeseus-ai/RiskClassifier}}
}

Contact

For inquiries, please reach out to [email protected] or visit LinkedIn.