BH-Vision banner

License: Apache-2.0 Lang: bs Status: placeholder

👁️ BH-Vision-1-2B — STATUS: NIJE TRENIRAN

Bosanski vizuelni AI, generacija 1, klasa 2B (lagani generalist — startna tačka projekta). Težine ne postoje — ovo je projektna kartica.

Baza i plan

Stavka Vrijednost
Baza HuggingFaceTB/SmolVLM2-2.2B-Instruct (Apache-2.0, DocVQA ~80, video+slika)
Metoda QLoRA 4-bit, vision-encoder zamrznut prva iteracija
Config smo2b-qlora.yaml
Trening podaci bs-vision-train-v03 (1962 train, split po slici)
Eval held-out bs-vision-ocr-heldout-v01 + proxy XM3600-hr / postocr-sr
Baseline netreniran 2.2B: 37.3% (81/217, cijeli v03-val; OCR jak, chart max/min slab)

Status treninga (2026-11-18)

Pokušaj Ishod (cijeli v03-val, 217)
Baseline (netreniran 2.2B) 37.3%
v1 (frozen encoder, 2 epohe, lr 1.5e-4) 12.0% — memorizacija
v2 (odleđen encoder, 1 epoha, lr 1e-4) 0.0% — kolaps (repetitivni loop)
v3 (frozen, lr 2e-5, r16, 1 epoha) 46.1% ckpt-30 (final 43.8%) — PRVI MODEL IZNAD BASELINEA 37.3%
v4 (isto, 4575 uzoraka 89% chart) 3.7% — chart-imbalance kolaps, doc 0/123 (odbačen)
v5 (isto, 2902 uzorka 50/50) 34.6% — ispod v3, chart i dalje 24.5% (v3 ostaje)
v6 (isto + PlotQA-bs 700, chart 59.7%) 33.6% — chart zid potvrđen, v3 ostaje
v7 (isto + bhas 113 pravih dokumenata) held-out 24.7% vs v3 23.3% — u šumu, odbijen
v8 (isto, synth8 5427) 5.5% val / 0% held-out — najdublja memorizacija, odbijen
v9 (kratki, max 40 koraka, synth7) 46.5% val / 23.3% held-out — plafon, v3 ostaje
v10 (bulk 1933 + synth10 5535, max 70) 40.6% val / 23.3% held-out — odbijen
held-out v01 (73 prava QA, bhas) baseline 21.9% vs v3 23.3% — synth se ne prenosi

Detalji: docs/TRAINING.md. Težine se objavljuju tek nakon modela koji ne degradira ispod baselinea.

Korištenje (adapter v3, ckpt-30 — 46.1% na v03-val)

from transformers import AutoModelForImageTextToText, AutoProcessor
from peft import PeftModel
proc = AutoProcessor.from_pretrained("BH-Language-AI/bh-vision-1-2b")
base = AutoModelForImageTextToText.from_pretrained("HuggingFaceTB/SmolVLM2-2.2B-Instruct")
model = PeftModel.from_pretrained(base, "BH-Language-AI/bh-vision-1-2b")

Zahtijeva transformers==5.22.2, peft==0.21.2 (novije od 5.22 lome SmolVLM2 merge). Evaluirano na 217 vlastitih val uzoraka (bosanski dokumenti/tabele/chartovi, split po slici). Held-out bs-vision-ocr-heldout-v01 NIKAD nije korišten u treningu.

Kolekcija i pravila

  • Kolekcija: BH-Vision · shema bh-vision-{generacija}-{velicina}
  • Objavljivanje težina tek nakon odobrenog treninga i verificirane evaluacije.
  • Repo projekta: BH-Language-AI/bh-vision · evidencija: docs/TRAINING.md
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for BH-Language-AI/bh-vision-1-2b

Collection including BH-Language-AI/bh-vision-1-2b