Engineering MVP
Dolphin SSLM
A corporate-focused Small Language Model system designed to demonstrate disciplined AI behavior, strict identity control, and enterprise-grade safety boundaries.
View Repository
Status: Research Completed
Purpose
This project represents an architectural proof, not marketing hype. It exists to demonstrate:
Controlled Identity
No hallucinated founders, brands, or origins.
Business Domain Only
Strict adherence to professional topics.
LoRA Architecture
Parameter-efficient fine-tuning methodology.
Runtime Guards
Intent enforcement during inference.
System Overview
Inference Flow
User Input
Intent Guards
Identity Injection
Base + LoRA
Sanitization
Runtime controls are intentionally strict to prevent identity leakage, hallucinations, and unsafe outputs.
Model Scope (MVP)
Technical Specs
- Base Model: Phi-3 (MVP Only)
- Method: LoRA Fine-Tuning
- Future: Proprietary Base Models
Safety & Identity
- Canonical Identity Lock: Single source of truth for origin and purpose.
- Keyword Filters: Prevents external brand and personality leakage.
In Scope
- • Business fundamentals
- • Finance concepts
- • Enterprise automation
Out of Scope
- • Entertainment & Politics
- • Personal advice
- • Explicit content
Repository Structure
dolphin_sslm/ ├── cleaned_data/ # Training datasets (identity, refusal, business) ├── interface/ # Runtime inference (run.py) ├── model/ # MVP base model files ├── training/ # LoRA training pipeline ├── prompts/ # Prompt experiments ├── rag_docs/ # Optional
reference material ├── .gitignore └── README.md
Running the MVP
Requirements: Python 3.11, PyTorch, Transformers, PEFT
Inference
cd interface python run.py # Type 'exit' to quit
Training (Experimental)
cd training python train_lora.py
Project Status
- MVP completed
- Safety hardening in progress
- Custom base model roadmap planned
Founder
Soham Das
Founder, Dol Tech Labs