Porodica modela za odlučivanje na bosanskoj ijekavici + engleskom
3 veličine (2B/4B/9B) · jedan format podataka · zajedničke izlazne sheme · zajednički benchmark
Nosioci: Kolektiv d.o.o. + Gamatrade d.o.o. · Kontakt: xnetba@gmail.com
Primarni README na ijekavici: README.bs.md
Šta Skor Gdje Stock Tacit-4B, readout 16/19 (84%) held-out v1 v1.0 (mix v4) + readout, q2/q3 16/19 (84%) — paritet held-out v1 v1.0 generacija (tolerantno) 6/19 (32%) held-out v1 Format (strogi JSON) 18–20/20 (mix v9: 18/20) held-out v1 Jev 1.13 Free, DI staza 14/14 (100%), p50 256 ms 18 scenarija A/B Sudzuki-4b-cloud, DI staza 15/18 (83%) held-out v1 Proizvodna konfiguracija: v0.4-rc + AnyJev readout = 17/19, prvi put iznad stocka. Pravilo repoa: bez fabriciranja — svaka brojka ispod je iz stvarnog runa.
- FAZE 1–4 — gotove (research, arhitektura, dataseti/benchmark, infrastruktura).
- FAZA 5 — Linija 2 do v1.0:
BH-Language-AI/bh-decision-2-4b(LoRA na Tacit-4B; readout paritet 15–16/19; generacija tolerantno 7/19). SFT serija v1–v9: format popravljen podacima (0.05→0.90), prior mican (o2→o3→nijedan), rasuđivanje ostalo na stock nivou. Nalazi:docs/training/. - DI staza — Core/MCP/Skill/Plugin gotovi (71 test zeleno); sudzuki-4b-cloud profilisan (15/18, ECE 0.16); provider-selection čeka cijenu/licencu. Vlastiti pipeline ovdje ne radi bez ključa (Jev mrtav, Clef nema).
| Verzija | Readout | Generacija (tolerantno) | Generacija (strogo) | val_loss |
|---|---|---|---|---|
| stock Tacit-4B | 16/19 (84%) | 7/19 (37%) | — | — |
| ✅ v1.0 (mix v4, q2/q3) | 16/19 (84%) | 6/19 (32%) | 1/20 | 0.100 |
| v1.0 (mix v4, q2/q3) | 16/19 (84%) | 6/19 (32%) | 1/20 | 0.100 |
| v0.2-rc (mix v4) | 15/19 (79%) | 6/19 (32%) | 1/20 | 0.100 |
| v0.1 (mix v2) | — | 2/10 (20%) | 10/20 | 0.124 |
Eval set: bs-decision-heldout-v01 (zaključan, leakage 0). Slike/grafike: bedževi su shields.io, graf je mjereni held-out (matplotlib).
📈 Decision Intelligence benchmark (18 scenarija, 9 bs + 9 en) — klikni za detalje
| Ruta | Točnost (12.11.) | Rerun (13.11.) | p50 rerun | Reproducibilnost |
|---|---|---|---|---|
| A Jev-1.13-Free (hosted, $0) | 14/14 (100%) | 14/14 | 256 ms | bit-za-bit ✓ |
| B AnyJev Decider L0 (Tacit-4B, CPU) | 11/14 (79%) | 11/14 strogo | 17,1 s | 18/18 identično ✓ |
| C AnyJev Tacit L0 (CPU) | 11/14 norm. | 9/14 strogo (= 11 norm.) | 6,4 s | 18/18 identično ✓ |
| D Tacit Adaptive | 11/14 | nije ponavljan (D==C dokazano, cot 0/18) | — | — |
| E Clef lokalno | — | blokirano (treba GPU/download) | — | — |
Sirovo: benchmarks/reports/live-{A-jev,B-decider-l0,C-tacit-l0,D-tacit-adaptive}.json
*-rerun-20261113.json. Metodologija:docs/benchmarking/anyjev-comparison.md.
Agent s skillom je brži i bolji: L0 je jedan forward (CPU) ili jedan hosted poziv (p50 256 ms);
svaka odluka nosi kalibrirano povjerenje; ispod praga 0.80 slijedi eskalacija.
Mermaid dijagram toka: docs/architecture/system-architecture.md.
| Sloj | Broj | Komanda |
|---|---|---|
| Python (pytest) | 24 zeleno | python3 -m pytest tests/ -q |
| TypeScript (core/mcp/plugin/integracija) | 47 zeleno | node --experimental-strip-types --test tests/core/normalize.test.ts tests/core/providers.test.ts tests/core/anyjev.test.ts tests/mcp/server.test.ts tests/plugin/loading.test.ts tests/integration/project.test.ts |
pip install pytest pyyaml
python3 -m pytest tests/ -q
python3 scripts/train.py --config configs/2b.yaml # dry-run: samo validira i ispisuje plan
python3 scripts/evaluate.py --ref ref.json --pred pred.json [--out out.json] # +Brier scoringconfigs/ trening konfiguracije 2b/4b/9b
scripts/ train.py, evaluate.py (+Brier), build_bh2_mix*.py, readout_eval.py, pilot.py
slurm/ kaggle/ LUMI sbatch + Kaggle kerneli (mix v2–v9, svi runovi)
tests/ 24 pytest + 47 TS (core/mcp/plugin/integracija)
docs/ model-selection, architecture, decisions, training (v1–v9 + readout izvještaji), providers (jev/clef/sudzuki)
benchmarks/ harness (accuracy + Brier) + reports
datasets/ train-mix-v1/v2, bh-decision-2 (mix v2–v9), expanded-v05..v09 human, heldout-v1 (zaključan), external (CSQA+QASC+ETHICS)
packages/ decision-core, decision-mcp, opencode-plugin
skills/ decision-intelligence (SKILL.md)
Skill + MCP server + OpenCode plugin + zajednički Core. Ne mijenja trening kod: packages/, skills/, config/.
- Arhitektura:
docs/architecture/system-architecture.md,docs/architecture/decision-protocol.md - Provideri:
docs/providers/jev.md,docs/providers/clef-flash.md,docs/providers/sudzuki.md,docs/decisions/provider-selection.md - Instalacija:
docs/installation/opencode.md,docs/installation/local-clef.md - Verzije: basic (Jev Free) vs advanced (lokalni Tacit-4B) —
docs/installation/tiers.md - Privatnost:
docs/security/privacy.md— Benchmark:benchmarks/reports/phase-6.md
Apache License 2.0 — vidi LICENSE. Doprinosi: CONTRIBUTING.md. Upravljanje: GOVERNANCE.md. Sigurnost: SECURITY.md.
Vidi CITATION.cff.
