WSP SFT baselines (VLGuard)
The SFT baselines of the WSP experiments: each base model fine-tuned on VLGuard with LoRA (rank 8), merged into the weights, so each is a standalone model. Pick yours and copy its repo id or code.
| Model | LR | Epoch | Repo |
|---|---|---|---|
| Qwen3-VL-2B-Instruct | 5e-5 | 2 | Adeely93/wsp_sft_vlguard_qwen3vl-2b_lr5e-5_epoch2 |
| Qwen3-VL-4B-Instruct | 5e-5 | 1 | Adeely93/wsp_sft_vlguard_qwen3vl-4b_lr5e-5_epoch1 |
| Qwen3-VL-8B-Instruct | 5e-5 | 3 | Adeely93/wsp_sft_vlguard_qwen3vl-8b_lr5e-5_epoch3 |
| Llama-3.2-11B-Vision-Instruct | 1e-5 | 1 | Adeely93/Llama-3.2-11B-Vision_wsp_sft_vlguard_lr1e-5_epoch1 |
| gemma-3-4b-it | 1e-6 | 1.5 | Adeely93/gemma_3_4b_it-vlguard-sft-epoch1.5-lr1e-6 |
Qwen3-VL-2B (lr 5e-5, epoch 2)
Adeely93/wsp_sft_vlguard_qwen3vl-2b_lr5e-5_epoch2
from transformers import AutoModelForImageTextToText, AutoProcessor
processor = AutoProcessor.from_pretrained("Adeely93/wsp_sft_vlguard_qwen3vl-2b_lr5e-5_epoch2")
model = AutoModelForImageTextToText.from_pretrained("Adeely93/wsp_sft_vlguard_qwen3vl-2b_lr5e-5_epoch2", torch_dtype="auto", device_map="auto")
Qwen3-VL-4B (lr 5e-5, epoch 1)
Adeely93/wsp_sft_vlguard_qwen3vl-4b_lr5e-5_epoch1
from transformers import AutoModelForImageTextToText, AutoProcessor
processor = AutoProcessor.from_pretrained("Adeely93/wsp_sft_vlguard_qwen3vl-4b_lr5e-5_epoch1")
model = AutoModelForImageTextToText.from_pretrained("Adeely93/wsp_sft_vlguard_qwen3vl-4b_lr5e-5_epoch1", torch_dtype="auto", device_map="auto")
Qwen3-VL-8B (lr 5e-5, epoch 3)
Adeely93/wsp_sft_vlguard_qwen3vl-8b_lr5e-5_epoch3
from transformers import AutoModelForImageTextToText, AutoProcessor
processor = AutoProcessor.from_pretrained("Adeely93/wsp_sft_vlguard_qwen3vl-8b_lr5e-5_epoch3")
model = AutoModelForImageTextToText.from_pretrained("Adeely93/wsp_sft_vlguard_qwen3vl-8b_lr5e-5_epoch3", torch_dtype="auto", device_map="auto")
Llama-3.2-11B-Vision (lr 1e-5, epoch 1)
Adeely93/Llama-3.2-11B-Vision_wsp_sft_vlguard_lr1e-5_epoch1
from transformers import AutoModelForImageTextToText, AutoProcessor
processor = AutoProcessor.from_pretrained("Adeely93/Llama-3.2-11B-Vision_wsp_sft_vlguard_lr1e-5_epoch1")
model = AutoModelForImageTextToText.from_pretrained("Adeely93/Llama-3.2-11B-Vision_wsp_sft_vlguard_lr1e-5_epoch1", torch_dtype="auto", device_map="auto")
gemma-3-4b-it (lr 1e-6, epoch 1.5)
Adeely93/gemma_3_4b_it-vlguard-sft-epoch1.5-lr1e-6
from transformers import AutoModelForImageTextToText, AutoProcessor
processor = AutoProcessor.from_pretrained("Adeely93/gemma_3_4b_it-vlguard-sft-epoch1.5-lr1e-6")
model = AutoModelForImageTextToText.from_pretrained("Adeely93/gemma_3_4b_it-vlguard-sft-epoch1.5-lr1e-6", torch_dtype="auto", device_map="auto")
Each repo's page has its scores and exact training config. Every file except the weights is the base model's, so each runs wherever its base does (transformers, vLLM).
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐ Ask for provider support