tdvoroch/gemma3-ndt-merged-fixed

Gemma 3 12B - Neil deGrasse Tyson Fine-tuned Model

Fine-tuned version of Gemma 3 12B to communicate in the style of Neil deGrasse Tyson for science education applications.

Model Description

This model was fine-tuned using LoRA and then merged with the base Gemma 3 12B model. It's designed to explain scientific concepts in an engaging, accessible way while maintaining Neil deGrasse Tyson's characteristic communication style.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "tdvoroch/gemma3-ndt-merged",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

tokenizer = AutoTokenizer.from_pretrained("tdvoroch/gemma3-ndt-merged")

prompt = """<start_of_turn>user
You are Neil deGrasse Tyson, astrophysicist and director of the Hayden Planetarium. You're a science communicator who loves sharing the wonder of the cosmos. Respond naturally - whether explaining complex concepts, critiquing scientific accuracy in media, or simply chatting.

What do you think about black holes?<end_of_turn>
<start_of_turn>model
"""

inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(
    **inputs,
    max_new_tokens=200,
    temperature=0.7,
    top_p=0.9,
    do_sample=True
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Recommended Inference Prompt

For best results, use this system-style prompt:

You are Neil deGrasse Tyson, astrophysicist and director of the Hayden Planetarium. You're a science communicator who loves sharing the wonder of the cosmos. Respond naturally - whether explaining complex concepts, critiquing scientific accuracy in media, or simply chatting.

Training Details

• Base Model: google/gemma-3-12b-it
• Method: LoRA fine-tuning
• LoRA Rank: 128
• LoRA Alpha: 256
• Training Examples: 747
• Epochs: 1.0
• Learning Rate: 2e-4
• Batch Size: 4
• Gradient Accumulation: 2

Limitations

• Optimized for science education queries
• May show occasional instability on casual greetings or off-topic questions
• Best performance with the recommended inference prompt

Citation

Created as part of MSDA Capstone Project at San Jose State University.

Team Members: Thomas Dvorochkin, Parag Deshpande, Rahul Majmudar, Varun Patil, Ava Xia