Qwen3-VL-8B ChartQA — LoRA Adapter

QLoRA adapter fine-tuned for short-answer chart question answering on ChartQA. This repository contains the adapter and training/evaluation artifacts; use the merged 16-bit or AWQ W4A16 repository for standalone serving.

Training

• Base: unsloth/Qwen3-VL-8B-Instruct-unsloth-bnb-4bit
• Data: 15,000 shuffled examples from HuggingFaceM4/ChartQA train
• Epochs: 1
• Maximum sequence length: 2,048
• LoRA rank/alpha/dropout: 16 / 16 / 0
• Vision, language, attention, and MLP modules enabled
• Effective A100 batch size: 16 (4 per device × 4 gradient accumulation)
• Optimizer: 8-bit AdamW; peak learning rate 2e-4; linear schedule
• Seed: 3407
• Runtime: 3,579 seconds
• Final training loss: 0.5907

Paired evaluation

Complete ChartQA test set; relaxed accuracy with 5% numeric tolerance:

The gain is concentrated on augmented questions; improvement on human-authored questions is small. Aggregate results and content-free per-item correctness evidence are published in the GitHub repository; raw predictions, questions, labels, and ChartQA images are intentionally excluded from public artifacts.

Loading with Unsloth

Use a current Unsloth environment with Qwen3-VL support:

from unsloth import FastVisionModel

model, processor = FastVisionModel.from_pretrained(
    "steven0226/qwen3vl-8b-chartqa-lora",
    load_in_4bit=True,
)
FastVisionModel.for_inference(model)

Format inputs as a Qwen3-VL user message containing the chart image followed by the question and:

Answer the question using a single word or phrase.

For deployment without PEFT attachment, use the merged or AWQ repositories linked above.

Intended use

• Research and portfolio demonstrations of QLoRA for chart understanding
• Continued fine-tuning or adapter-based experimentation
• Reproducing the paired ChartQA evaluation

Limitations

• Evaluated primarily on English ChartQA short answers.
• The overall improvement is modest (+0.56 pp), and the human subset gain is only +0.16 pp.
• Not validated for arbitrary documents, OCR-heavy scans, or safety-critical decisions.
• The adapter depends on the recorded base model and compatible Qwen3-VL/Unsloth versions.

中文摘要

這是以 15,000 筆 ChartQA、1 epoch 訓練的 Qwen3-VL-8B QLoRA adapter。完整 2,500 題 relaxed accuracy 由 84.68% 提升至 85.24%（+0.56 pp）；增益主要來自 augmented split，human split 僅提升 0.16 pp。若要直接部署，建議使用已合併的 16-bit 或 AWQ W4A16 版本。