| --- |
| title: MainStack |
| emoji: 🛠️ |
| colorFrom: blue |
| colorTo: indigo |
| sdk: static |
| pinned: false |
| --- |
| |
| # MainStack |
|
|
| ### ServiceNow Agentic Delivery — and the open models behind it. |
|
|
| [mainstack.co.uk](https://www.mainstack.co.uk/) · [LinkedIn](https://linkedin.com/company/mainstack-it) |
|
|
| MainStack is a consultancy specializing in ServiceNow Agentic Delivery. We build the AI that does the drafting work of a delivery engagement — and we release it openly. **marvy-1-14B** is our first public model: a fine-tuned LLM that drafts artifacts across the entire ServiceNow delivery lifecycle, from business analysis through Solution Design Documents, user stories, test cases, and validation. It's a first-draft specialist built for the people who do the work — solution architects, business analysts, technical consultants, project managers, and delivery leads. |
|
|
| ## 📊 Measured impact |
|
|
| On a **project- and customer-disjoint** held-out test set, `marvy-1-14B` cuts |
| perplexity on real delivery artifacts by **32% overall vs. the unmodified base |
| model** — and by **75–86%** on structured artifacts like systems inventories, |
| requirements, and stakeholder registers. Same weights, adapter on vs. off; the |
| difference is the fine-tune. See the charts on the |
| [model card](https://huggingface.co/MainStack/marvy-1-14B#evaluation). |
|
|
| ## 🔒 Built responsibly |
|
|
| marvy was trained on real engagement artifacts, which means privacy was a design constraint, not an afterthought. Every training artifact was rigorously anonymized and redacted to **zero residual PII**, verified by an automated leakage scanner. Evaluation was run on a **project- and customer-disjoint** held-out split, so the reported perplexity reflects genuine generalization to unseen work — not memorization. We frame marvy honestly: it accelerates first drafts; it does not replace the judgment of a consultant, and it is not a tool-use or agentic fine-tune. |
|
|
| ## 📦 Available formats |
|
|
| | Repo | Format | Use case | |
| |------|--------|----------| |
| | [MainStack/marvy-1-14B](https://huggingface.co/MainStack/marvy-1-14B) | Merged FP16 | Full-precision inference and serving | |
| | [MainStack/marvy-1-14B-lora](https://huggingface.co/MainStack/marvy-1-14B-lora) | LoRA adapter | Compose on top of Qwen2.5-14B-Instruct | |
| | [MainStack/marvy-1-14B-GGUF](https://huggingface.co/MainStack/marvy-1-14B-GGUF) | GGUF quants | Local runs on Apple Silicon, LM Studio, Ollama | |
|
|
| ```bash |
| ollama run hf.co/MainStack/marvy-1-14B-GGUF:Q4_K_M |
| ``` |
|
|
| ## About |
|
|
| MainStack specializes in ServiceNow Agentic Delivery, building open, fine-tuned models that draft the artifacts of a delivery engagement — and releasing them to the practitioners who use them. Learn more at [mainstack.co.uk](https://www.mainstack.co.uk/). |
|
|
| *Released under Apache-2.0. Built with Qwen.* |
|
|
