Nicesse-RARA-3B is a complete 3B chart visual question answering model in the Qwen2.5-VL family. It is trained with Reweighted Answer-Restricted Adaptation (RARA), which combines challenge-aware example exposure with answer-focused supervision for numerical and comparative chart reasoning. The repository contains the complete model checkpoint and the matching processor configuration for direct use with transformers.

Model Details

Item	Value
Model family	Qwen2.5-VL-3B
Parameter scale	3B
Architecture	Vision-language generation model
Training objective	Challenge-aware sampling with answer-restricted supervision
Primary use	Chart visual question answering and numerical reasoning
Inputs	Chart image and natural-language question
Outputs	Short free-form answer

Benchmark Results

The table reports publicly available results on four chart VQA benchmarks. Scores are dataset-level accuracy in percent; higher is better. ChartQAPro additionally has a strict-audit score of 22.28.

Method	ChartQA	PlotQA	EvoChart	ChartQAPro	Average
Qwen2.5-VL-3B (direct prompting)	82.00	80.50	48.72	25.70	59.23
Qwen2.5-VL-3B (CoT)	73.12	52.72	29.60	15.80	42.81
ChartGemma	76.44	33.28	36.96	10.93	39.40
SketchVL-3B	77.20	48.32	47.28	44.15	54.24
Qwen2.5-VL-3B SFT	83.08	74.18	46.08	23.56	56.73
Qwen2.5-VL-3B DPO	75.42	53.86	34.80	15.95	45.01
Qwen2.5-VL-3B (A+F+L)	76.72	56.22	38.88	17.55	47.34
Qwen2.5-VL-3B (A+F+L+Tasks)	81.80	76.24	51.68	27.66	59.35
Chart-RVR-3B	84.56	78.68	53.36	28.38	61.25
Chart-RVR-3B-Hard	85.76	77.90	54.24	28.64	61.64
Qwen2.5-VL + Nicesse-RARA-3B	83.76	89.42	54.32	37.94	66.36

The Nicesse-RARA-3B configuration was selected using a frozen development split. It was then evaluated once on four frozen benchmark manifests. The result is strongest on PlotQA among the listed comparable public 3B models; its ChartQAPro result follows the published-code-compatible evaluation protocol.

Quick Start

python

from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration

model_id = "Junlaii/Nicesse-RARA-3B"
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    model_id, torch_dtype="auto", device_map="auto"
)
processor = AutoProcessor.from_pretrained(model_id)

messages = [{
    "role": "user",
    "content": [
        {"type": "image", "image": "path/to/chart.png"},
        {"type": "text", "text": "What is the value of the blue bar in 2010?"},
    ],
}]
prompt = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = processor(text=[prompt], images=["path/to/chart.png"], return_tensors="pt").to(model.device)
generated_ids = model.generate(**inputs, max_new_tokens=768)
print(processor.batch_decode(generated_ids, skip_special_tokens=True)[0])

Intended Use and Limitations

Nicesse-RARA-3B is intended for research on chart understanding, visual question answering, and numerical reasoning. It can make visual-reading, arithmetic, formatting, and ambiguous-question errors. Benchmark results do not establish reliability for high-stakes decisions; validate outputs independently in any downstream deployment.

License and Attribution

This model is distributed under the included Qwen Research License Agreement. It is improved using Qwen. Please follow the license terms and cite the relevant Qwen2.5-VL and RARA work when appropriate.