Gemma4-31B-IT-ChartQA-Adapted – AI Model by Arnav-Gr0ver | AlphaNeural AI

You can deploy this model and start earning money today!

Arnav-Gr0ver/
Gemma4-31B-IT-ChartQA-Adapted

like

0

adapter-transformers

Arnav-Grover/Adapted_ChartQA_Visual_Primitives

HuggingFaceM4/Chart.QA

Views

No views yet

Gemma-4-31B-IT Chart Visual Primitives

An SFT LoRA fine-tune of google/gemma-4-31B-it-VLM trained to ground chart visual reasoning in coordinate primitives (points and polygons).

Overview

Developed by: Arnav Grover (Purdue University)
Base Model: google/gemma-4-31B-it-VLM
License: MIT
Task: Visual Question Answering / Chart QA
Dataset: 4,968 grounded visual reasoning traces derived from ChartQA

Model Performance

Evaluated on 500 test questions from ChartQA:

Model / Configuration	Numeric QA (%)	Other QA (%)	Overall (%)
Gemma 4 31B-IT (Direct)	73.5	88.9	79.6
Gemma 4 31B-IT (Reasoning Chain)	80.1	88.4	83.4
Adapted (Ours)	85.8	94.9	89.4

Training Details

Method: SFT via LoRA
LoRA Configuration: $r = 16$, $\alpha = 32$, Dropout = 0
Target Modules: q_proj, k_proj, v_proj, o_proj
Learning Rate: 1e-4 (Cosine schedule, 0.1 warmup ratio)
Epochs / Steps: 2 epochs / 86 steps
Hardware: 1 x NVIDIA A100 (80GB)

Quickstart & Usage

python

import torch
from transformers import AutoProcessor, AutoModelForCausalLM

model_id = "google/gemma-4-31B-it-VLM"
adapter_id = "your-username/adaption_chart_qa_numeric"

processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
model.load_adapter(adapter_id)

# Inference standard ChartQA input