Titan-Hohmann

🚀 Pixtral 12B Fine-Tuned on Titan-Hohmann-Transfer-Orbit (v1.0)

🔢 Version 1.0, built on mistralai/Pixtral-12B-Base-2409.

🌟 Overview

Fine-tuned variant of Pixtral 12B for orbital mechanics with emphasis on Hohmann transfer orbits. Supports multimodal (image + text) inputs and text outputs.

🔧 Model Details

• Base: mistralai/Pixtral-12B-Base-2409
• Version: 1.0
• Type: 🖼️ Multimodal (Vision + Text)
• Params: ~12B (decoder) + ~400M vision encoder
• Languages: 🇺🇸 English
• License: 📄 MIT (base model Apache 2.0, attribution retained)

🎯 Intended Use

• 🛰️ Hohmann transfer ∆v estimation
• ⏱️ Transfer-time approximations
• 🔍 Orbit analysis aids and reasoning

🚫 Out of Scope

• 🧭 Mission design or trajectory planning without independent verification
• 🛠️ Flight software or any system where an error has physical consequences
• 🌐 Domains outside orbital mechanics

🚀 Quickstart

🌐 vLLM (multimodal)

from vllm import LLM
from vllm.sampling_params import SamplingParams

llm = LLM(model="Taylor658/Titan-Hohmann", tokenizer_mode="mistral")
sampling = SamplingParams(max_tokens=512, temperature=0.2)

messages = [
    {
        "role": "user",
        "content": [
            {"type": "text", "text": "Given this diagram, estimate the delta-v for a Hohmann transfer to Titan."},
            {"type": "image_url", "image_url": {"url": "https://example.com/orbit_diagram.png"}}
        ]
    }
]
resp = llm.chat(messages, sampling_params=sampling)
print(resp[0].outputs[0].text)

🤗 Transformers (text-only demo)

from transformers import LlavaForConditionalGeneration, AutoProcessor
import torch

model_id = "Taylor658/Titan-Hohmann"
processor = AutoProcessor.from_pretrained(model_id)
model = LlavaForConditionalGeneration.from_pretrained(model_id, torch_dtype="auto", device_map="auto")

prompt = "Compute approximate delta-v for a Hohmann transfer to Titan. State assumptions."
inputs = processor(text=prompt, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=512, temperature=0.2)
print(processor.decode(out[0], skip_special_tokens=True))

📊 Training Data

• Dataset: Taylor658/titan-hohmann-transfer-orbit
• Modalities: 📝 text (explanations), 💻 code (snippets), 🖼️ images (orbital diagrams)
• Provenance: synthetic; no flight or ephemeris data

📏 Evaluation

Quantitative benchmarks are not yet published for v1.0. Evaluation to date is manual review of sampled outputs against closed-form Hohmann transfer calculations. Results will be added in a later version.

⚠️ Limitations

• 🔢 v1.0 is a first release; behavior will change in later versions
• 🎯 Optimized for Hohmann transfers and related reasoning; may degrade elsewhere
• 🔬 Trained on synthetic data; ∆v and transfer-time outputs are approximations, not mission values
• 📏 No published benchmarks yet
• 👁️ Vision limitations of the base model carry over (small text, dense diagrams)
• 💾 Requires sufficient GPU VRAM for best throughput

🙏 Acknowledgements

• Base model by Mistral AI (Pixtral 12B, Apache 2.0)
• Dataset by A Taylor

📞 Contact Information

• Author: 👨‍🚀 A Taylor
• Support: 💬 Hugging Face discussions tab on this repository
• Companion physics library: 🔗 https://github.com/ATaylorAerospace/HohmannHET

---