Powell-Phi3-Mini — Jerome Powell Style Language Model

🎯Summary

Powell-Phi3-Mini is an fine-tuned language model that replicates Federal Reserve Chair Jerome Powell's distinctive communication style, tone, and strategic hedging patterns. This project showcases expertise in modern LLM fine-tuning techniques, parameter-efficient training methods, and responsible AI development — demonstrating industry-ready machine learning engineering skills.

---

🚀 Key Features & Capabilities

Style Mimicry & Linguistic Analysis

• ✅ Authentic Communication Style: Replicates Powell's cautious, data-dependent rhetoric
• ✅ Strategic Hedging Patterns: Maintains appropriate uncertainty in speculative scenarios
• ✅ Domain-Specific Responses: Handles economic and monetary policy discussions contextually
• ✅ Refusal Training: Appropriately declines to provide financial advice or policy predictions (to an extent)

Technical Implementation

• ✅ Efficient Architecture: Built on Microsoft Phi-3-mini-4k-instruct (3.8B parameters)
• ✅ Scalable Training: LoRA r=16, alpha=32 configuration optimized for consumer GPUs
• ✅ Deployment Flexibility: Available as lightweight adapter or full merged model
• ✅ Integration Ready: One-line inference with Hugging Face Transformers

---

💻 Implementation Examples

Production Ready - Merged Model

from transformers import AutoTokenizer, AutoModelForCausalLM

# One-line model loading
tokenizer = AutoTokenizer.from_pretrained("BoostedJonP/powell-phi3-mini")
model = AutoModelForCausalLM.from_pretrained("BoostedJonP/powell-phi3-mini", device_map="auto")

# Economic analysis prompt
prompt = "How is the current labor market affecting your inflation outlook?"
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
response = model.generate(**inputs, max_new_tokens=200, do_sample=True)
print(tokenizer.decode(response[0], skip_special_tokens=True))

---

📊 Technical Specifications & Training Pipeline

Model Architecture

Component	Specification
Base Model	microsoft/Phi-3-mini-4k-instruct (3.8B parameters)
License	MIT License (Commercial Use Approved)
Fine-tuning Method	QLoRA with PEFT integration
Context Length	4,096 tokens
Training Hardware	NVIDIA TESLA P100 (16GB VRAM)

Training Configuration

Hyperparameter	Value	Rationale
LoRA Rank (r)	16	Optimal parameter/performance balance
LoRA Alpha	32	2x rank for stable training
Dropout Rate	0.05	Regularization without overfitting
Learning Rate	1.5e-4	Conservative rate for stable convergence
Scheduler	Cosine decay	Smooth learning rate reduction
Training Epochs	3	Prevents overfitting on specialized domain
Sequence Length	1,536 tokens	Optimized for dataset
Precision	Mixed fp16	2x memory efficiency, maintained accuracy

Dataset & Methodology

• Data Source: Public domain FOMC transcripts and Federal Reserve speeches -> Jerome Powell Press Release Q&A
• Data Processing: Instruction-response pairs extracted from press conferences -> Jerome Powell Press Release SFT data processing
• Available on HuggingFace
• Quality Control: Manual review and filtering for authentic Powell communication patterns

---

📈 Performance Metrics & Evaluation

Quantitative Results

Metric	Baseline (Phi-3)	Powell-Phi3-Mini	Improvement
Powell-style Classification	NA	NA	NA
Economic Domain Accuracy	NA	NA	NA
Response Coherence (BLEU)	NA	NA	NA

Qualitative Assessment

• NA

---

🌐 Deployment & Access

🚀 Live Demo

Try Powell-Phi3-Mini Interactive Demo →

📦 Model Downloads

• Adapter Version: BoostedJonP/powell-phi3-mini-adapter
• Merged Model: BoostedJonP/powell-phi3-mini (Full Model - 7.4GB)

🔗 Resources

• GitHub Repository: Complete training code and evaluation scripts
• Technical Blog Post: Detailed implementation walkthrough
• Hugging Face Collection: All model variants and datasets

---

⚖️ Responsible AI & Legal Compliance

Ethical Considerations

• ⚠️ No Official Affiliation: Not endorsed by or affiliated with the Federal Reserve System
• ⚠️ Educational Purpose Only: Designed for research, education, and demonstration purposes
• ⚠️ No Financial Advice: Model responses should not be interpreted as investment guidance
• ⚠️ Transparency: All training data sourced from public domain government transcripts

Licensing & Usage Rights

• Base Model License: MIT License (Microsoft Phi-3)
• Fine-tuned Weights: MIT License (Commercial use permitted)
• Training Data: Public domain (U.S. government works)
• Usage: Unrestricted for research, education, and commercial applications

---

👨‍💻 Connect & Collaborate

• GitHub: Jonathan Paserman
• Kaggle: Jonathan Paserman
• HuggingFace: Jonathan Paserman