Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more
Container

Run this model inference with full control and performance in your environment.

Learn more

Get help setting up a custom Dedicated Endpoints.

Talk with our engineer to get a quote for reserved GPU instances with discounts.

README

Данные

Использован открытый датасет deepvk/GQA-ru из коллекции DeepVK VLM на Hugging Face. Данные использовались как VQA: изображение, вопрос на русском языке и короткий эталонный ответ.

В smoke-обучении использовано 1 000 train и 100 validation примеров из подготовленных локальных манифестов GQA-ru.

Обучение

ПараметрЗначение
Base modelQwen/Qwen2.5-VL-3B-Instruct
AdapterLoRA
Target modulesq_proj, k_proj, v_proj, o_proj
Rank / alpha / dropout16 / 32 / 0.05
Epochs1.0
Batch size / grad accumulation1 / 16
Learning rate2e-4
Precisionbf16
Seed42

Training metrics:

MetricValue
train_loss0.6548236324673608
eval_loss0.47339919209480286
train_runtime_sec714.4626
train_samples_per_second1.4
eval_samples_per_second7.061

Benchmark

Официальный benchmark smoke выполнен через lmms-eval на задаче gqa-ru:

МодельSamplesExactMatch
Qwen/Qwen2.5-VL-3B-Instruct1000.39
LoRA adapter1000.48

Улучшение: +0.09 ExactMatch absolute, +23.1% relative.

Ограничение: использован lmms-eval --limit 100, поэтому это bounded official-task smoke, а не полный leaderboard score.

Использование

python

from peft import PeftModel
from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration
base = "Qwen/Qwen2.5-VL-3B-Instruct"
adapter = "lockR/vk-vlm-gqa-ru-qwen25vl-3b-lora-smoke"
processor = AutoProcessor.from_pretrained(base, trust_remote_code=True)
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(base, trust_remote_code=True)
model = PeftModel.from_pretrained(model, adapter)

Репозиторий проекта

https://github.com/L0ckR/VK_education_vllm

Model provider

lockR

Model tree

Base

Qwen/Qwen2.5-VL-3B-Instruct

Adapter

this model

Modalities

Input

Text, Image

Output

Text

Pricing

Dedicated Endpoints

View details

Supported Functionality

Model APIs

Dedicated Endpoints

Container

More information

Explore FriendliAI today