Technical Approach
How HuskyBot is being built.
A modern parameter-efficient fine-tuning approach + retrieval-augmented generation, all on top of an open-weight foundation — designed to scale.
The Approach
Open-weight foundation. Academic specialization. Lean by design.
We are starting from a balanced, open-weight base model — either Mistral 7B or Llama 3 8B — and using parameter-efficient fine-tuning (LoRA / QLoRA) to specialize it for academic work. This is the same technique that lets startups ship competitive AI without billion-dollar budgets.
Mistral 7B / Llama 3 8B
Open-weight base model. Strong general capabilities, well-supported by the community.
LoRA / QLoRA
Parameter-efficient fine-tuning — trains only small adapter layers, runs on a single GPU.
RAG Architecture
Retrieval over academic corpus grounds answers in real sources — not memorized text.
Training Strategy
Layered. From general to specific.
Three layers of fine-tuning, each building on the last — starting broad, ending tailored to SCSU.
Foundation
Open-weight base model + academic writing style guides (APA, MLA, Chicago, IEEE).
Domain Expansion
General academic knowledge — broad scholarly literature across disciplines.
Specialization
Specific fields and SCSU-relevant domains based on usage patterns.
Data Sources
The academic corpus, free and open.
Our Phase 1 training data comes entirely from open-access academic sources — giving us multidisciplinary breadth without licensing entanglements.
arXiv
STEM preprints
Bulk S3Semantic Scholar / S2ORC
Multidisciplinary
50–100M papersPubMed Central
Biomedical
Open Access SubsetOpenAlex
Citation networks
CC0 metadataCORE
OA aggregator
Worldwide reposStyle Guides
Citation manuals
APA • MLA • Chicago • IEEEPhase 2 will explore licensed sources like ACM Digital Library, IEEE Xplore, and Elsevier ScienceDirect TDM — with negotiated addenda for LLM training. Recent legal precedent (Bartz v. Anthropic) suggests training on lawfully-acquired content is likely Fair Use.
Training Pipeline
Six steps from raw papers to deployed bot.
Data Harvest
Pull from arXiv, Semantic Scholar, PubMed Central, OpenAlex, and academic style guides.
Clean & Format
Filter, deduplicate, format into instruction-tuning pairs (Q&A, paraphrase, citation).
Fine-Tune (LoRA/QLoRA)
Parameter-efficient adaptation of the base model on academic writing tasks.
Wrap with RAG
Layer in retrieval over academic corpus for grounded, citable answers.
Evaluate
Automated metrics + human eval on citation accuracy, helpfulness, and tone.
Package & Deploy
Publish to web app + library embed; ship downloadable GGUF for offline use.
Output Formats
Three ways to use HuskyBot.
GGUF (offline)
Quantized model file for local use with Ollama or llama.cpp — runs on consumer hardware.
Hugging Face Hub
Standard ML format for the broader research community — open weights, model card, transparent benchmarks.
Web API Endpoint
Hosted endpoint for the SCSU library website embed and standalone web app — always up to date.
Built lean. Designed to scale.
The same architecture that serves a single tester scales to thousands of concurrent users. The same model that runs on your laptop also runs on a campus inference server. That is the power of an open-weight, open-source approach — we are not boxed in by anyone.
See Budget Allocation