Skip to content

About

Reproducibilan framework za fine-tuning malih bosanskojezicnih modela sa function calling i agentic ponasanjem (AgentMujo).

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

agentmujo-training

Typing SVG

Stars License Top lang Last commit matrix


AgentMujo Training Framework

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

CI License

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.

Razumijevanje za 5 minuta

  1. 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.
  2. Thinking/non-thinking = JEDAN model. enable_thinking zastavica chat templatea prebacuje režim; 4 deployment profila dijele iste težine.
  3. Alati: configs/tools.yaml je Single Source of Truth (14 alata: service_*, disk/memory/cpu, process_list, network_status, port_check, package_*, file_*, terminal kao fallback). Model MORA preferirati high-level alat nad sirovom terminal komandom.
  4. Sigurnost: MODEL → TOOL → POLICY ENGINE → LINUX. Fine-tuning nije zaštita; allow / confirmation_required / deny odlučuje izvršavanje.
  5. Linija: baza → Function-Calling SFT → Agentic SFT → Bench → najbolji adapter → Q8 → 4 profila.

Brzi start

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 testova

Struktura

  • configs/ — project, models (pinovane revizije), datasets, tools, training
  • src/agentmujo_training/ — cli (amj), tools, datasets, benchmark, context, registry, policy (Policy Engine + Executor), training/, evaluation/
  • schemas/ — dataset, tool, experiment, model JSON Scheme
  • datasets/canonical/ — 1252 function-calling + 541 agentic-terminal + 1118 bosnian-core + 45 thinking + DPO parovi; sirovi podaci na HF Hubu
  • datasets/splits/ — zamrznuti train/valid/test + leakage_report.json
  • benchmark/ — AgentMujo-Bench (16 kategorija, 48 slučajeva v0.5) + agent_eval.py + format-probe + bosanska proba
  • quantization/ — GGUF Q8 pipeline (convert --no-mtp + block_count patch protokol)
  • workers/kaggle/ — T4 worker (SFT + DPO + smoke), notebook generator, 40 jobova
  • deployment/ — Ollama Modelfile profili (thinking/non-thinking, zajednički GGUF)
  • docs/ — PROJECT_SPEC, TEST_PLAN, HF_PUBLISHING
  • tests/ — registry, validator (+safety), benchmark

Pravila

  • 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

Objavljeno na Hugging Face Hubu

  • 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

part of xnet-ba matrix — open-source AI iz BiH

About

Reproducibilan framework za fine-tuning malih bosanskojezicnih modela sa function calling i agentic ponasanjem (AgentMujo).

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages