SkyAsl/Nanbeige4.1-VLM-LoRA

Nanbeige4.1-VLM — Stage 2 LoRA Weights

LoRA adapter weights from Stage 2 instruction fine-tuning. Use this if you want the smallest download or want to merge manually.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

base = AutoModelForCausalLM.from_pretrained(
    "Nanbeige/Nanbeige4.1-3B",
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
)
model = PeftModel.from_pretrained(base, "SkyAsl/Nanbeige4.1-VLM-Stage2-LoRA")

# For full VLM inference, use the merged version instead:
# SkyAsl/Nanbeige4.1-VLM-Stage2

LoRA Config

• r: 64
• alpha: 128
• target: q_proj, k_proj, v_proj, o_proj
• Dataset: LLaVA-Instruct-150K
• Base: SkyAsl/Nanbeige4.1-VLM-Base (Stage 1)