TBAC-VLR1-3B-SFT

Overview

This is a multimodal language model fine-tuned by Tencent PCG Basic Algorithm Center. Based on Qwen2.5-VL-3B-Instruct, TBAC-VLR1-3B-SFT undergoes SFT training using 40k sft data filtered from OpenR1-Math-220k. TBAC-VLR1-3B then employs GRPO (Group Relative Policy Optimization) and adapts Clip-Higher from DAPO, achieving state-of-the-art results on several multimodal reasoning benchmarks among models of the same size.

Performance

Model	Average	MathVista	MathVision	MathVerse	DynaMath	LogicVista
Qwen2-VL-2B	22.4	48.0	16.1	17.5	3.8	26.6
InternVL2.5-2B	23.8	51.1	14.0	22.3	4.4	27.3
InternVL3-2B	31.5	57.6	20.2	24.5	14.8	40.3
Qwen2.5-VL-3B	33.6	61.2	21.9	31.2	13.2	40.3
VLM-R1-3B-Math-0305	34.1	62.7	21.9	32.2	13.0	40.5
Taichu-VLR-3B	34.3	64.9	23.1	32.1	12.6	38.7
VLAA-Thinker-Qwen2.5VL-3B	35.7	61.0	24.4	36.4	18.2	38.5
TBAC-VLR1-3B-preview	36.3	64.8	25.0	33.2	17.7	40.8
TBAC-VLR1-3B-SFT	35.3	57.0	27.4	41.1	15.0	36.1
TBAC-VLR1-3B	36.7	57.5	28.7	41.1	16.1	40.0

The results of our model are self-reported, obtained by running evaluations offline on each benchmark.

Usage

from transformers import Qwen2_5_VLForConditionalGeneration, AutoTokenizer, AutoProcessor
from qwen_vl_utils import process_vision_info

model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    "TencentBAC/TBAC-VLR1-3B-SFT", torch_dtype="auto", device_map="auto"
)

processor = AutoProcessor.from_pretrained("TencentBAC/TBAC-VLR1-3B-SFT")

messages = [
    {
        "role": "system",
        "content": "You FIRST think about the reasoning process as an internal monologue and then provide the final answer. The reasoning process MUST BE enclosed within <think> </think> tags. The final answer MUST BE put in \\boxed{}."
    },
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "image": image_path,
            },
            {"type": "text", "text": query},
        ],
    }
]

# Preparation for inference
text = processor.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)
image_inputs, video_inputs = process_vision_info(messages)
inputs = processor(
    text=[text],
    images=image_inputs,
    videos=video_inputs,
    padding=True,
    return_tensors="pt",
)
inputs = inputs.to("cuda")

# Inference: Generation of the output
generated_ids = model.generate(**inputs, max_new_tokens=128, do_sample=False)
generated_ids_trimmed = [
    out_ids[len(in_ids) :] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
output_text = processor.batch_decode(
    generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
)
print(output_text)

Citation

If you find our model useful in your research, please consider giving ❤️ and citations. Thanks!

@misc{Ou2025TBACVLR1,
  title = {TBAC-VLR1-3B},
  author = {Ou, Linyu and Xu, Junzhe and Yin, Yuyang},
  year = {2025},
  url = {https://huggingface.co/TencentBAC/TBAC-VLR1-3B},
}

About