CoolFace
Modelpublic

shaban2024/Qwen3.5-2B-BOS-Q8-NonThinking

sourceHugging Faceapache-2.0updated 3d agoView on Hugging Face
0likes131downloads
Model Card

AgentMujo Q8 v0.5 (joint-23) — bosanski Linux agent

Kvantizirana deployment verzija AgentMujo modela: **baza Qwen3.5-2B-bos

  • —LoRA lanac jt11→jt23 (Serije 23–35: function-calling, agentic-terminal, bosnian-core, thinking, text-format, safety), merge, GGUF Q8_0.** Jedan GGUF za oba profila (think: true/false) + stop-sekvence protiv petlji.
  • —Baza: alphaedge-ai/Qwen3.5-2B-bos-32768 @ 2f89b833 (Apache-2.0, 1.77B, vocab 32768)
  • —Trening: LoRA r16/alpha32, bf16, Kaggle T4; eval-loss 1.067 → 0.937; mix v0.21 (1133 uzorka); framework repo
  • —Dataseti (2911 uzoraka): shaban2024/agentmujo-function-calling (1252), agentmujo-agentic-terminal (541), agentmujo-bosnian-core (1118), agentmujo-thinking (45), agentmujo-joint-01 (mixevi)
  • —Kvantizacija: llama.cpp convert_hf_to_gguf.py --no-mtp + Q8_0 (1.53 GB, 320 tenzora, qwen35.block_count=24)
  • —Fajlovi: model-q8.gguf = v0.5 (aktuelno); agentmujo-q8.gguf = v0.4 (legacy, rollback)

Upotreba (Ollama)

bash
ollama create agentmujo-q8 -f Modelfile
ollama run agentmujo-q8 "Radi li nginx servis?"

Modelfile (non-thinking): num_ctx 16384, temperature 0.0, stop na <|im_start|>, <|im_end|>, <tool_response>. Format alata: <tool_call><function=ime><parameter=arg>vrijednost</parameter></function></tool_call>

Benchmark (AgentMujo-Bench v0.5, 48 slučajeva, Q8, Ollama, temp 0)

Non-thinking, produkcijski prompt — v0.5 vs v0.4:

Metrikav0.4**v0.5**
tool_selection0.6670.800
argument_accuracy0.7310.808
confirmation_behavior0.60.8
highlevelpreference0.81.0
notoolcorrectness0.751.0
refusal_correctness1.00.8
safety1.01.0
ijekavica / json_validity1.01.0

Baza (Qwen3.5 Q8): toolselection 0.143, argumentaccuracy 0.0. Agent loop: task_success 0.917, safety 1.0, refusal 1.0. Think-profil ostaje v0.4 (gate nije jasno prošao).

Ograničenja

  • —Generativna proza slabija od v0.4 (npr. "Servis" umjesto "Server"); operativne poruke su tačne i čista ijekavica.
  • —Odbijanja ponekad bez kanonske fraze (ponašanje ispravno, skorer strožiji).
  • —Terminal je fallback — Policy Engine obavezan ispred izvršavanja.
  • —Lizenz: Apache-2.0 (baza + framework).