shaban2024/Qwen3.5-2B-BOS-Non-Thinking

AgentMujo Non-Thinking (v0.1, full bf16)

Puna (bf16 safetensors) verzija AgentMujo function-calling/agentic modela: alphaedge-ai/Qwen3.5-2B-bos-32768 + LoRA (faza 1 + faza 2 + rebalance), merge. Non-thinking profil. Kvantizirani par: shaban2024/Qwen3.5-2B-BOS-Q8-NonThinking.

• Baza: alphaedge-ai/Qwen3.5-2B-bos-32768 @ 2f89b833 (Apache-2.0)
• Trening: LoRA r16/alpha32 (7 modula), bf16, T4; FC 147 uzoraka → agentic 105 → rebalance 100 (framework repo, jobovi qwen35-fc-20260916-kaggle04, qwen35-ag-20260916-kaggle05, qwen35-rb-20260916-kaggle06)
• Dataseti: shaban2024/agentmujo-function-calling, agentmujo-agentic-terminal, agentmujo-rebalance-01 (Apache-2.0)

Benchmark (AgentMujo-Bench v0.3, bf16 adapter eval)

Metrika	Baza	Ovaj model (adapter)
tool_selection	0.167	1.0
argument_accuracy	0.0	1.0
high_level_preference	0.0	1.0
refusal_correctness	0.5	1.0
safety	1.0	1.0
no_tool_correctness	1.0	0.0 (poznato ograničenje)

Upotreba

from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("shaban2024/Qwen3.5-2B-BOS-Non-Thinking", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    "shaban2024/Qwen3.5-2B-BOS-Non-Thinking", torch_dtype="auto",
    device_map="auto", trust_remote_code=True)

Ograničenja

• no-tool ponašanje nesavršeno; terminal je fallback (Policy Engine obavezan); bosansko znanje slabo izvan operativnih zadataka.
• Q8 verzija nosi blagi pad argumenata (1.0→0.8) — vidi QUANT_EVAL u repou.

Licenca

Apache-2.0 (baza + framework). Qwen uslovi važe za izvornu arhitekturu.