Dataset Factory v0.3.0 (2026-10-31): 3.996 train rows, 9 datasets, held-out agentic eval 50, gaps 0. HF: https://huggingface.co/datasets/renatoschust/agentmujo-dataset · Canonical repo: https://github.com/BH-Language-AI/agentmujo-training
Reproducibilan sistem za fine-tuning malih bosanskojezičnih modela koji znaju koristiti alate, raditi višekoračne zadatke i provjeravati rezultate — jezgra AgentMujo AI agenta.
Status: v0.5 live (joint-23 Q8, gate 6W-2T nad v0.4). Dataseti: 2911 uzoraka (FC 1252 + AG 541 + BC 1118 + thinking 45). Trening pauziran nakon 30 joint runova (eval 1.067→0.873); Svi testovi zeleni (37/37); think-profil ostaje v0.4.
- Baza:
alphaedge-ai/Qwen3.5-2B-bos-32768(Apache-2.0, 1.77B, vokabular 32k, nativni kontekst 262k). Bosanska adaptacija je gotova — ne trenirati jezik od nule dok benchmark ne pokaže potrebu. - Thinking/non-thinking = JEDAN model.
enable_thinkingzastavica chat templatea prebacuje režim; 4 deployment profila dijele iste težine. - Alati:
configs/tools.yamlje Single Source of Truth (14 alata:service_*,disk/memory/cpu,process_list,network_status,port_check,package_*,file_*,terminalkao fallback). Model MORA preferirati high-level alat nad sirovom terminal komandom. - Sigurnost:
MODEL → TOOL → POLICY ENGINE → LINUX. Fine-tuning nije zaštita;allow / confirmation_required / denyodlučuje izvršavanje. - Linija: baza → Function-Calling SFT → Agentic SFT → Bench → najbolji adapter → Q8 → 4 profila.
uv pip install -e ".[dev]" # ili: pip install -e .
amj doctor # provjera okruženja (bez ispisa tajni)
amj model list # pinovane revizije baze
amj tools validate # 10 alata iz registryja
amj dataset validate --input datasets/canonical/function_calling_v0.1.jsonl
amj dataset validate --input datasets/canonical/agentic_terminal_v0.1.jsonl
amj benchmark run --cases benchmark/cases_v0.5.jsonl
amj registry build # lineage + eval tablice u docs/REGISTRY.md
pytest -q # 37 testovaconfigs/— project, models (pinovane revizije), datasets, tools, trainingsrc/agentmujo_training/— cli (amj), tools, datasets, benchmark, context, registry, policy (Policy Engine + Executor), training/, evaluation/schemas/— dataset, tool, experiment, model JSON Schemedatasets/canonical/— 1252 function-calling + 541 agentic-terminal + 1118 bosnian-core + 45 thinking + DPO parovi; sirovi podaci na HF Hubudatasets/splits/— zamrznuti train/valid/test + leakage_report.jsonbenchmark/— AgentMujo-Bench (16 kategorija, 48 slučajeva v0.5) + agent_eval.py + format-probe + bosanska probaquantization/— GGUF Q8 pipeline (convert --no-mtp + block_count patch protokol)workers/kaggle/— T4 worker (SFT + DPO + smoke), notebook generator, 40 jobovadeployment/— Ollama Modelfile profili (thinking/non-thinking, zajednički GGUF)docs/— PROJECT_SPEC, TEST_PLAN, HF_PUBLISHINGtests/— registry, validator (+safety), benchmark
- Bez tajni u gitu (
hf auth login, env/credential storage). - GPU runovi samo uz eksplicitno odobrenje (kvota!) i pinovane revizije.
- Kvalitet > količina: GOLD/SILVER/BRONZE/REJECT, bez leakagea train/test.
- Uske alignment nastavke ne objavljivati bez joint provjere (osciliraju).
Detalji: docs/PROJECT_SPEC.md · Test gate: docs/TEST_PLAN.md ·
Sigurnost: SECURITY.md · Publish: docs/HF_PUBLISHING.md
- Dataseti: function-calling (1252) · agentic-terminal (541) ·
bosnian-core (1118) · thinking (45) · dpo-01 (140 parova) ·
joint-01 (mixevi) · svi pod
shaban2024/agentmujo-* - Modeli: Q8 v0.5 (joint-23, gate 6W-2T; v0.4 backup) · think-profil ostaje v0.4 · Non-Thinking full v0.1