Qwen3-VL-2B-ChartQA-SFT

Fine-tuned Qwen3-VL-2B-Thinking on the ChartQA dataset for advanced chart understanding, data extraction, and visual reasoning.

Model Description

This model is a specialized version of Qwen3-VL-2B-Thinking, fine-tuned using LoRA (Low-Rank Adaptation). It is designed to perform "System 2" thinking—generating step-by-step reasoning traces before providing the final answer—making the decision process regarding chart interpretation transparent and verifiable.

Key Features

• 🧠 Explicit Reasoning: Uses <think></think> tags to articulate the visual analysis process (e.g., identifying axes, comparing bars) before concluding.
• 📊 Chart Specialization: Optimized for bar charts, line graphs, and pie charts found in the ChartQA dataset.
• ⚡ Efficient: Fine-tuned with LoRA (~2% trainable parameters) while retaining the base model's general capabilities.
• 🎯 Structured Output: Trained to return answers in a strict JSON format {"answer": "value"} for easy programmatic parsing.

Training Details

Base Model

• Model: Qwen/Qwen3-VL-2B-Thinking
• Parameters: 2B
• Architecture: Vision-Language Model with built-in Chain-of-Thought (CoT) capability.

Fine-tuning Configuration

• Method: LoRA (Low-Rank Adaptation)
• Rank (r): 16
• Alpha: 32
• Target Modules: q_proj, k_proj, v_proj, o_proj
• Trainable Parameters: ~2%

Training Hyperparameters

• Epochs: 1
• Batch Size: 1 (with Gradient Accumulation = 8)
• Learning Rate: 2e-5
• Optimizer: AdamW
• Scheduler: Cosine
• Precision: FP16
• Max Sequence Length: 2048

Usage

Installation

pip install git+[https://github.com/huggingface/transformers](https://github.com/huggingface/transformers) # Install latest for Qwen3 support
pip install torch pillow