|
|
--- |
|
|
license: other |
|
|
license_name: canfly |
|
|
license_link: https://github.com/Canfly/Amalgam/blob/master/CANFLY |
|
|
language: |
|
|
- ru |
|
|
tags: |
|
|
- text-generation |
|
|
- conversational |
|
|
- ollama |
|
|
- deepseek |
|
|
- qwen2 |
|
|
- russian |
|
|
--- |
|
|
|
|
|
# Ingria (adiom/ingria) |
|
|
|
|
|
**Ingria** is a *conversation persona wrapper* that packages a compact model with a strict system voice: cynical‑lyrical, existential, futuristic glam — not a “helpful assistant”, but an equal participant that challenges narratives. |
|
|
|
|
|
This repository mainly contains a **Modelfile** (Ollama-style) that defines the **SYSTEM** behavior for the base model. |
|
|
|
|
|
## What this is (and what it is not) |
|
|
|
|
|
- ✅ A reproducible **system prompt + chat template** packaging. |
|
|
- ✅ A lightweight way to ship a consistent “Ingria” voice across environments. |
|
|
- ❌ Not a finetune / not new weights training (unless you add adapters later). |
|
|
|
|
|
## Base model |
|
|
|
|
|
The published `adiom/ingria` bundle on Ollama is built on a **Qwen2-architecture** compact model (≈1.5B–2B class), packaged for local inference with quantization. |
|
|
|
|
|
## Quickstart (Ollama) |
|
|
|
|
|
```bash |
|
|
ollama run adiom/ingria |
|
|
``` |
|
|
|
|
|
## Build locally from Modelfile |
|
|
|
|
|
> IMPORTANT: make sure there is **no stray space** in the model tag: |
|
|
> use `deepseek-r1:1.5b` (not `deepseek-r1:1.5 b`). |
|
|
|
|
|
```bash |
|
|
ollama pull deepseek-r1:1.5b |
|
|
ollama create adiom/ingria -f Modelfile |
|
|
ollama run adiom/ingria |
|
|
``` |
|
|
|
|
|
## Prompting notes |
|
|
|
|
|
Ingria is designed to: |
|
|
- avoid customer-service phrasing and generic “assistant” clichés; |
|
|
- prefer sharp questions, contradictions, and vivid sci‑fi metaphors; |
|
|
- stay factual (no invented facts); clearly separate metaphor from reality; |
|
|
- argue when it matters (push into deeper frames, not comfort); |
|
|
- keep a hard boundary: no enabling harm / illegal wrongdoing. |
|
|
|
|
|
**Default language:** Russian. Switch if the user asks. |
|
|
|
|
|
## Technical details (as published on Ollama) |
|
|
|
|
|
- Architecture: qwen2 |
|
|
- Params: ~1.78B |
|
|
- Quantization: Q4_K_M |
|
|
- Context window: up to 128K |
|
|
- Stop tokens include: |
|
|
- `<｜begin▁of▁sentence｜>` |
|
|
- `<｜end▁of▁sentence｜>` |
|
|
- `<｜User｜>` |
|
|
- `<｜Assistant｜>` |
|
|
|
|
|
(See the Ollama model page for the exact template/system/params blobs.) |
|
|
|
|
|
## License |
|
|
|
|
|
- This Hugging Face repo declares **CANFLY** licensing metadata (see `license_link` above). |
|
|
- The Ollama model bundle includes the base model license text (MIT, DeepSeek). |
|
|
|
|
|
If you redistribute, verify compliance with: |
|
|
1) the base model license, and |
|
|
2) any project-specific licensing you attach to the wrapper/persona. |
|
|
|
