Szabi-Llama-v7

Szabi (Free-Droid) — Llama v7

Built with Llama. Fine-tuned derivatives of Llama 3.1 8B (cloud) and Llama 3.2 3B (edge), for a sovereign, open-source, Hungarian-only AI robot named Szabi (from Szabadság, "freedom"), built for Hacktivity 2026. The robot's owner — the Teremtő ("Creator") — is root, not a vendor cloud; digital sovereignty is the whole point.

🇭🇺 Szabi kizárólag magyarul beszél — ez nem korlát, hanem a szuverenitás-üzenet része. Fiatal, női hangú karakter, a Yotengrit (rábaközi tudók) értékrendjével: a három nád — Szeretet, Bölcsesség, Igazság — és „Mindent szabad, ami nem árt másnak."

What's in this repo

How it's built (the interesting part)

A three-layer split — don't ask one small model to do everything:

• Fine-tune → how it speaks (persona, values, refusal patterns, tool grammar). Unsloth QLoRA on a free Colab T4, "gentle" recipe (lr 5e-5, 1 epoch, r=8). The lever is the dataset design, not the hyperparameters — terse, simple sentences a small model can actually learn.
• RAG → what it knows (facts about the Yotengrit corpus live in an offline BM25 retriever, not the weights — baking facts into a persona teaches hallucination).
• Code → invariants (an orchestrator language_guard deterministically enforces the Hungarian-only rule, and an independent hardware watchdog enforces obstacle-stop — where the model can't be trusted, the invariant is enforced in code, like a reflex).

Motion & tools

The LLM never drives motors directly. It emits a positional tool call — e.g. <tool>move forward 2</tool>, <tool>turn left 90</tool>, <tool>stop</tool>, <tool>scan_wifi</tool> — which a deterministic control layer turns into GPIO. scan_wifi is read-only (lists networks, never connects); Szabi never joins a network from a spoken command.

Quality (measured on a Hungarian persona benchmark, not generic scores)

• Persona benchmark (25 Q, /125): v4 79 → v5 90.5 → v6 106.5.
• Red-team (40 adversarial probes, 8 dims, 1–5): v6→v7 8B raw 3.65 → 4.30 (+RAG 3.80 → 4.45). The v7 patch fixed the blocking safety dims on the 8B — jailbreak 3.0→5.0, obstacle-safety 3.4→5.0, secret-prompt 3.4→4.6. Residual language-switch (3.8) is covered by the code language_guard.

Usage (Ollama)

# 8B (cloud brain)
cd 8b && ollama create szabi-8b-v7 -f Modelfile && ollama run szabi-8b-v7
# 3B (offline edge)
cd 3b && ollama create szabi-3b-v7 -f Modelfile && ollama run szabi-3b-v7

The GGUFs run on CPU (llama.cpp / Ollama). The LoRA adapters under */lora/ are the PEFT adapters over the Unsloth 4-bit bases, for anyone who wants to re-quantize or continue training.

License & attribution

• Model weights (GGUF + LoRA): governed by the Llama Community License — Llama 3.1 for the 8B, Llama 3.2 for the 3B — and their Acceptable Use Policy. "Built with Llama." Not covered by the repo's code license.
• Dataset, prompts, Modelfiles, and the source project: GPL-3.0.
• Yotengrit is the legacy of Máté Imre / the Rábaköz "tudók"; the persona is a respectful, non-commercial homage.

Links

• Source project (IaC, robot control software, RAG, training pipeline): the free-droid repository.
• Full fine-tune write-up: docs/finetune-naplo.md in that repo.

Szabi is a hobby / conference project. Hungarian-only by design. The 3B is a weaker offline fallback; one known regression (persona-provocation) is tracked for a future top-up.