internvl3_5-1b-balanced
OpenGVLab/InternVL3_5-1B-Pretrained 를 한국어 멀티모달 데이터로 파인튜닝한 InternVL3.5 specialist.
Hyperparameters
- num_train_epochs:
1
- steps:
5162 / max 5162
- train_batch_size:
8
- peak learning_rate:
9.999984252764371e-05
Training Loss
- init 1.4746 → final 0.9904 (min 0.9242)
| step | loss |
|---|
| 10 | 1.4746 |
| 520 | 1.0770 |
| 1030 | 1.0217 |
| 1540 | 1.0139 |
| 2050 | 0.9842 |
| 2560 | 0.9938 |
| 3070 | 0.9712 |
| 3580 | 0.9627 |
| 4090 | 0.9475 |
| 4600 | 0.9431 |
| 5110 | 0.9660 |
| 5160 | 0.9904 |
Training Data
구성: 40개 서브셋 (한국어 specialist SFT)
| subset | repeat |
|---|
hf_korLlava_Caption | 2.7 |
kr_recap_caption | 1.2857 |
chart2table_en | 0.0475 |
chartRqa1 | 0.3167 |
chartRqa2 | 0.475 |
dvqa_en | 0.0475 |
en_chartqa_chart | 0.5278 |
en_figureqa_chart | 0.475 |
en_mapqa_chart | 1.3571 |
kisti_documen_Reason | 2.25 |
en_imgtext_doc | 0.5625 |
en_vwi_doc | 0.9 |
out_kor_llava | 1.9501 |
en_allava_general | 3.0 |
aihub_mathMultiple | 0.8449 |
aihub_mathSubjective | 1.7883 |
en_geoqa_math | 1.0 |
en_hme_formula | 1.6667 |
en_iconqa_math | 0.6667 |
en_mavis_math | 0.3333 |
hf_latexUpdate | 0.6667 |
aihub_subjectTxt_OCR | 0.625 |
aihub_visual_OCR | 0.8333 |
kisti_arxiv_OCR | 0.8333 |
kopub_SDSKoPub_OCR | 0.3077 |
en_chrome_ocr | 1.4205 |
en_iam_ocr | 2.2093 |
en_llavar_ocr | 0.6316 |
en_ocrvqa_ocr | 0.25 |
aihub_visual_ShortQA | 0.7167 |
kisti_hanbat_Reason | 0.7167 |
kisti_hanbat_Vqa | 0.86 |
en_aokvqa_general | 1.3438 |
aihub_subjectImg_Parse | 1.35 |
en_tabmwp_table | 0.6136 |
en_tatqa_table_text | 1.0385 |
tableVqa_Caption | 0.675 |
tableVqa_Reason | 0.675 |
ko_alpaca_textonly | 0.9006 |
en_evol_textonly | 0.3152 |
Usage
from transformers import AutoModel, AutoTokenizer
import torch
m = AutoModel.from_pretrained("yujuyeon/internvl3_5-1b-balanced", torch_dtype=torch.bfloat16,
trust_remote_code=True).eval().cuda()
tok = AutoTokenizer.from_pretrained("yujuyeon/internvl3_5-1b-balanced", trust_remote_code=True, use_fast=False)