georgeanton/sifta-corvid-qwen35

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

Benchmark Results (from SIFTA head-to-head experiment)

The 2B won the head-to-head. It's faster, smaller, and has fewer RLHF scars. The 4B is not shipped.

Three-Layer Architecture

🦐 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

# 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`:

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:

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