jeremymoon/gemma-sentiment-nuclear

Gemma 7B Sentiment Analysis - Nuclear Energy

google/gemma-7b fully fine-tuned for 3-class tweet sentiment classification (negative / neutral / positive), with a focus on nuclear energy discourse. Trained by Jeremy Moon, AIMS Lab, University of Michigan.

Labels

0: negative, 1: neutral, 2: positive

Training

• Full fine-tune with bf16, batch size 24, learning rate 1e-5, 5 epochs, on 4x RTX 6000 Ada
• 829,252 training tweets and 207,313 validation tweets (general tweets plus high-confidence nuclear energy tweets)

Results

Test set	Accuracy	Macro-F1
sample (3,000 tweets)	98.43%	0.984
test (259,143 tweets)	97.37%	0.974

Usage

import torch
from transformers import pipeline
clf = pipeline("sentiment-analysis",
model="jeremymoon/gemma-sentiment-nuclear",
                torch_dtype=torch.bfloat16, device_map="auto")
clf("Nuclear power is key to decarbonizing the grid.")

Use of this model is subject to the Gemma Terms of Use.