| # Arriella Fleet Whitepaper | |
| **Scout · Growth · Flagship · Ascension — plus Grapevine multimodal & Model Interior Projection** | |
| *Version 4.1 · July 2026* | |
| **arXiv manuscripts (publication SoT):** [`../arxiv/`](../arxiv/) · submit via [`../arxiv/SUBMISSION.md`](../arxiv/SUBMISSION.md) → HF Papers after arXiv ID | |
| --- | |
| ## Read this first | |
| Arriella is a **local training factory** that turns open instruct foundations into named small LMs you own on disk, plus a **Model Interior Projection (MIP)** viewer that makes inference inspectable in 3D — attention trails on a real parameter cloud, not a black-box chat box. | |
| Organization: **Infinidev Corp** (Infinidev). Lead developers: **Beelzebub4888** and **Tcoder**. | |
| ### What problem we solve | |
| 1. Useful custom models on a consumer GPU (8 GB Blackwell, Windows) | |
| 2. **Role-scoped tiers** for field and business use — not one generic chatbot | |
| 3. Continuous improvement via **eat** (weight ingest / distill) and **grow** (post-deploy LoRA) | |
| 4. Stability knobs: enterprise base blend, capability/math/thinking anchors, gates before Ollama | |
| 5. A product nobody else ships: **live interior visualization of decode** | |
| 6. A separate **native multimodal** extension when text routing is not enough | |
| ### What “Arriella” means | |
| | Sense | Meaning | | |
| |-------|---------| | |
| | Project | This repo — scripts, configs, `fleet/`, docs, MIP | | |
| | Persona | System identity in datasets + chat | | |
| | Dataset | `arriella_qa` (~90k Pile-derived Alpaca records) | | |
| | Core products | `arriella-scout`, `arriella-growth`, `arriella-flagship`, `arriella-ascension` | | |
| | Extension | `arriella-grapevine` (Omni multimodal — not a fifth core text model) | | |
| Arriella is **not** a Meta/Alibaba/DeepSeek foundation model. Foundations are clay; products are what comes out of the factory. | |
| --- | |
| ## Product thesis — stable models for fields & business | |
| Enterprise and field deployments fail when a single mid-size chat model is asked to be edge-fast, domain-deep, reason-heavy, and multimodal at once. Arriella ships **specialists with shared factory discipline**: | |
| | Need | Route to | Why | | |
| |------|----------|-----| | |
| | Low VRAM / high throughput | Scout | Smallest core text tier | | |
| | Domain QA / continuation | Growth | Trained on full Arriella QA pile | | |
| | General instruction ops | Flagship | Heretic + premium teacher distill | | |
| | Explicit step-by-step reasoning | Ascension | DeepSeek-R1 distill native think | | |
| | Image / audio in the loop | Grapevine | Qwen2.5-Omni-3B student | | |
| Reliability comes from **gates and anchors**, not slogans: capability / math / thinking / identity / fleet-spec datasets, optional enterprise base blend (~8–15% foundation retention), bakeoffs vs same-size stock instruct, and honest modality claims. | |
| --- | |
| ## Fleet tiers (Jul 2026 stance) | |
| ### Core text fleet (exactly four) | |
| | Tier | Size (Ollama) | Foundation | Training path | Product stance | | |
| |------|---------------|------------|---------------|----------------| | |
| | **Flagship** | ~1.5B | Qwen2.5-1.5B-Instruct | Heretic → premium QLoRA distill | Primary quality / general ops | | |
| | **Growth** | ~1.2B | Llama-3.2-1B-Instruct | LLaMA-Factory on `arriella_qa` | Primary mid / domain tier | | |
| | **Ascension** | ~1.8B | DeepSeek-R1-Distill-Qwen-1.5B | LLaMA-Factory + Flagship catch-up | Reasoning / native think — **not** automatically smarter than Flagship | | |
| | **Scout** | ~494M | Qwen2.5-0.5B-Instruct | Heretic → QLoRA | Present but **currently underpowered** — edge experiments, not demos | | |
| Canonical numbers and truth rules: [`fleet/CORE_FLEET_SPEC.md`](../../fleet/CORE_FLEET_SPEC.md). Benchmarks: [../benchmarks/README.md](../benchmarks/README.md). | |
| ### Multimodal extension (not core text) | |
| | Product | Foundation | Stance | | |
| |---------|------------|--------| | |
| | **Grapevine** | Qwen2.5-Omni-3B | Technical preview for vision/audio; modular runtime LoRA + mmproj. Inkling is a capability target only — no Inkling weights. | | |
| Guide: [../guides/grapevine.md](../guides/grapevine.md). HF draft: [grapevine-hf.md](grapevine-hf.md). | |
| --- | |
| ## Training framework | |
| ``` | |
| foundation HF weights | |
| ├─ Path A (Scout / Flagship): Heretic abliteration → QLoRA distill → merge | |
| ├─ Path B (Growth / Ascension): LLaMA-Factory LoRA → merge | |
| └─ Path C (Grapevine): Omni Thinker QLoRA → merge / runtime adapter → GGUF + mmproj | |
| ↓ | |
| fleet/<tier>/model/ (standalone HF) | fleet/inkling/ (Grapevine) | |
| ↓ | |
| enterprise base blend (optional) + capability / fleet-spec anchors | |
| ↓ | |
| optional eat (weight ingest / distill) + post-deploy grow | |
| ↓ | |
| GGUF → ollama run arriella-* | |
| ``` | |
| **Heretic** abliterates refusal directions (not fine-tuning). Distillation then teaches what to say under QLoRA. | |
| **Pufferfish** in this repo is the **long-run policy / WebUI preset layer** (`fleet_pufferfish.py`, `arriella_integration/`), not the upstream LightLM trainer as the primary engine. | |
| Orchestrator: `scripts/train_fleet.py` + `fleet/fleet_manifest.yaml`. | |
| --- | |
| ## Eat & grow | |
| After merge, capabilities improve by: | |
| - **Weight ingest** — same-arch tensor merge (linear / TIES / DARE-TIES / …) | |
| - **Distill grow** — cross-architecture teacher imitation (prefer larger teachers) | |
| - **Organic grow** — post-deploy LoRA on anchors / datasets | |
| - **Enterprise align** — base-preservation blend + capability re-eval (`fleet_enterprise_align.py`) | |
| Honest language SoT: [../guides/eat-system.md](../guides/eat-system.md). | |
| --- | |
| ## MIP — Model Interior Projection | |
| MIP loads a HF/Ollama-backed checkpoint, samples parameter space into ≤120k GPU display points, and during decode: | |
| - Paints prompt-specific attention / residual activity on the cloud | |
| - Traces a gold polyline (one node per generated token) | |
| - Forges a deterministic projected **tesseract** from that trail | |
| - Can **repeat** a prompt N times and **compare** solids (seed, tokens, geometry) | |
| Guide: [../guides/mip-viewer.md](../guides/mip-viewer.md). Entry: `scripts/fleet_mip_web.py`. | |
| This is the differentiator. Chat UIs show tokens. MIP shows **where** attention landed in the model’s own geometry. | |
| --- | |
| ## Multimodal (honest) | |
| **Core text tiers** do not ingest raw pixels on 8 GB. Images go Gemma3 → categorized text → fleet model. Thinking/vision behaviors are **ingested** via teacher traces. | |
| **Grapevine** is the native Omni path (text + image + audio in Ollama; video validated pre-export). Details: [multimodal.md](multimodal.md), [../guides/grapevine.md](../guides/grapevine.md). | |
| --- | |
| ## Competitive posture | |
| - Compete with **same-size open instruct** checkpoints via bakeoffs. | |
| - Do not claim frontier closed-model parity. | |
| - Prefer measured tables in `docs/benchmarks/` over slogans. | |
| - Prefer **role fit** (field/business job) over “one model to rule them all.” | |
| --- | |
| ## Publication (arXiv → Hugging Face Papers) | |
| - Manuscript set: [`../arxiv/`](../arxiv/) | |
| - Workflow: [`../arxiv/SUBMISSION.md`](../arxiv/SUBMISSION.md) | |
| - Hub guide: [`../guides/hugging-face-publish.md`](../guides/hugging-face-publish.md) | |
| Do **not** host papers as Spaces or model-only-markdown repos. Spaces are for runnable demos; Paper Pages require arXiv IDs. | |
| --- | |
| ## Document history | |
| - v2.5 — fleet narrative without MIP as product centerpiece (`docs/archive/2026-07/WHITEPAPER.md`) | |
| - v3.0 — reorganized docs; MIP as unique product; honest Scout status; three-tier only | |
| - v4.0 — four-core text fleet + Grapevine extension; field/business thesis; HF paper prep | |