HuskyBotSCSU Library AI
← Project Info

Timeline & Phases

From foundation to v1.0: May through November 2026.

Four phases. Integration, evaluation and launch now run September through November so testing gets the time it needs. Open-source academic AI, v1.0 target: November 30, 2026.

Where we are now

Status as of October 15, 2026

54%Overall progress · 13/24 items complete

Phase 1

Foundation & Outreach

In progress · 4/6

Phase 2

Training Pipeline & Model

In progress · 6/7

Phase 3

Integration & Evaluation

In progress · 2/6

Phase 4

Polish & v1.0 Launch

In progress · 1/5

Current focus: E4B training is finished and HuskyBot Lite can now be downloaded. Next: recruiting testers (anyone is welcome), before/after evaluation, and meetings on Repository access and licensed datasets.

DoneIn progressNot started
Phase 1In progress

Foundation & Outreach

May 15 – May 31, 2026

~2 weeks

Progress

  • Open-access papers harvested (OpenAlex, arXiv, DOAJ, Europe PMC) (Done)
  • GPU compute in place: RTX 5080 workstation + RunPod account (Done)
  • Outreach email drafts prepared for providers, vendors and SCSU (Done)
  • HuskyBot name, Circuit Husky logo and project website (Done)
  • SCSU Repository (OAI-PMH) access request with Library IT (In progress)
  • Licensed dataset agreements with vendors (Not started)

Key deliverables

Compute secured
Open datasets harvested
Outreach in motion
Phase 2In progress

Training Pipeline & Model

June 1 – August 31, 2026

~13 weeks

Progress

  • Base model chosen: Google Gemma 4 (E4B local, 31B cloud) with QLoRA (Done)
  • Training kit through hbt14: harvest, dataset, train, chat/eval, GGUF export, Windows installer (Done)
  • ~5,000-example academic dataset + 157 verified SCSU Library Q&A rows (Done)
  • Open-license filter and answer-only training (Done)
  • 30-step test run passed (loss 3.14 → 1.46) (Done)
  • Full E4B training run on the latest kit (Done)
  • 31B training run on RunPod (Not started)

Key deliverables

Reproducible training kit
Fine-tuned E4B adapter
31B model checkpoint
Phase 3In progress

Integration & Evaluation

September 1 – October 15, 2026

~6.5 weeks

Progress

  • Held-out eval sets: SCSU Library facts and academic tasks (Done)
  • No-fabrication examples: refuses to invent citations (Done)
  • Before/after evaluation of the fine-tuned model (Not started)
  • RAG over SCSU Library documents and research guides (Not started)
  • LibSearch / research guide integration (Not started)
  • Structured testing with library staff (Not started)

Key deliverables

Evaluation results
RAG pipeline
Library integrations
Phase 4In progress

Polish & v1.0 Launch

October 16 – November 30, 2026

~6.5 weeks

Progress

  • GGUF export and Windows installer for the downloadable model (Done)
  • Hugging Face model card and open-source repository (In progress)
  • Final polish on tone, safety and edge cases (Not started)
  • Soft launch on the SCSU Library website (Not started)
  • Launch communications with library leadership (Not started)

Key deliverables

v1.0 web release
Downloadable GGUF package
Open-source repository

Project start

May 15, 2026

Funds committed by

June 1, 2026

v1.0 target

November 30, 2026

Curious about how it works under the hood?

Read about the technical approach: fine-tuning strategy, RAG architecture, and data pipeline.

Technical Approach