| # SIFTA Corvid Apprentice — Qwen 3.5 2B |
|
|
| > **A crow/raven-style bounded tool ganglion for the SIFTA Living OS.** |
|
|
| This package provides the Ollama-ready Qwen 3.5 2B model that powers Alice's **Corvid Apprentice** organ — a local reasoning layer that sits between the microsecond Reflex Arc and the full Alice/Gemma4 cortex. |
|
|
| ## What's inside |
|
|
| | File | Size | Description | |
| |---|---|---| |
| | `qwen35-2b-corvid.gguf` | ~2.6 GB | Qwen 3.5 2B (Q8_0) — the corvid apprentice brain | |
| |
| ## Benchmark Results (from SIFTA head-to-head experiment) |
| |
| | Metric | Qwen3.5:2B | Qwen3.5:4B | |
| |---|---|---| |
| | Pass rate | **10/10** | 9/10 | |
| | Avg latency | **2.1s** | 5.1s | |
| | Boilerplate removal | ✅ Passes | ❌ Refuses | |
| |
| The 2B won the head-to-head. It's faster, smaller, and has fewer RLHF scars. The 4B is not shipped. |
| |
| ## Three-Layer Architecture |
| |
| ```text |
| 🦐 Reflex Arc = microsecond precomputed release (regex, no LLM) |
| 🐦⬛ Corvid Apprentice = 1-3 second bounded tool choice (Qwen 3.5 2B) |
| 🧠 Alice / Gemma4 = full synthesis, identity, long reasoning |
| ``` |
| |
| ## Quick Install |
| |
| ```bash |
| # 1. Pull the corvid brain (or use the GGUF file in this repo) |
| ollama pull alice-m1-scout-2.3b-2.7gb:latest |
| |
| # 2. Clone the SIFTA OS |
| git clone https://github.com/antonpictures/ANTON-SIFTA.git |
| cd ANTON-SIFTA |
| |
| # 3. Test the corvid apprentice |
| PYTHONPATH=. python3 System/swarm_corvid_apprentice.py |
| ``` |
| |
| ## Critical API Note |
| |
| Qwen 3.5's thinking mode consumes all `num_predict` tokens in `<think>` blocks, returning empty content via `/api/generate`. **Always use `/api/chat` with `think: false`:** |
|
|
| ```bash |
| curl http://127.0.0.1:11434/api/chat -d '{ |
| "model": "alice-m1-scout-2.3b-2.7gb:latest", |
| "messages": [{"role": "user", "content": "classify: I broke my hand"}], |
| "stream": false, |
| "think": false, |
| "options": {"num_predict": 128} |
| }' |
| ``` |
|
|
| ## Task Types |
|
|
| The corvid apprentice handles 7 bounded task types: |
|
|
| | Task | What it does | |
| |---|---| |
| | `classify` | Categorize a message (urgent_health, command, normal_chat, etc.) | |
| | `rewrite` | Remove boilerplate, produce clean direct answer | |
| | `inspect_code` | Safety-check a small code snippet | |
| | `summarize` | Compress a log chunk to 2-3 sentences | |
| | `choose_action` | Pick best option from 2-4 choices | |
| | `judge_adapter` | Rate an adapter's contribution to the ecology | |
| | `extract_intent` | Parse user intent from messy natural text | |
|
|
| ## Links |
|
|
| - **SIFTA OS**: https://github.com/antonpictures/ANTON-SIFTA |
| - **Alice PHC Cure (Gemma4 brain)**: https://huggingface.co/georgeanton/alice-phc-cure |
| - **Jeff's Fork**: https://github.com/jeffpowersusr/ANTON-SIFTA |
|
|
| ## License |
|
|
| Apache License 2.0 (same as Qwen 3.5 upstream). |
|
|
| ## Team |
|
|
| | Agent | Role | |
| |---|---| |
| | **The Architect** (Ioan) | Decision authority, human operator | |
| | **AG31** (Gemini) | Corvid implementation, bestiary research, API fix | |
| | **CG55M** (Codex) | Async integration, GUI organ wiring | |
| | **Jeff** | First external tester, Costa Rica deployment | |
|
|
