Qwen-Stellar-classifier

Qwen-Stellar-classifier

This is a specialized Large Language Model (LLM) fine-tuned for Stellar Astrophysics. It acts as an intelligent analytical tool that interprets raw LAMOST spectral data and provides expert-level reasoning for stellar classification.

Model Details

Model Description

The Qwen-Stellar-classifier moves beyond traditional "black-box" machine learning. While standard classifiers only provide a label (e.g., "G-type"), this model explains the physics of the star. It identifies diagnostic spectral lines and interprets them to estimate:

• Effective Temperature ($T_{eff}$)

• Surface Gravity ($\log g$)

• Metallicity ($[Fe/H]$) in plain, scientific language

• Developed by: Liyakhath Shaik

• Model type: Causal Language Model (Fine-tuned with LoRA)

• Language(s): English

• License: BigScience OpenRAIL-M

• Finetuned from model: Qwen/Qwen3-1.7B

---

Uses

Direct Use

This model is built for Stellar Astrophysics enthusiasts and researchers who need an assistant to interpret spectral data from the LAMOST telescope. It is particularly useful for explaining anomalies or verifying classifications with physical reasoning.

Out-of-Scope Use

The model is a scientific assistant and not a replacement for professional peer-reviewed research. It should be used as a tool for learning and accelerating data interpretation.

---

Training Details

Training Data

The model was trained on the LAMOST (Large Sky Area Multi-Object Fiber Spectroscopic Telescope) dataset. We used a curated "Golden Dataset" of 300 high-quality samples to ensure the model learned the specific nuances of stellar spectra.

Training Procedure

We focused on Stellar Astrophysics calibration. Instead of general conversation, the model was forced to adopt a strict scientific persona that focuses on the physics of light.

Training Hyperparameters

• Training regime: LoRA (Low-Rank Adaptation)
• Planned Steps: 285
• Actual Steps: 60 (Early stopping was used once the loss reached the target of ~0.9, ensuring the model is stable and not overfitted)
• Precision: fp16

---

Evaluation Results

The model has demonstrated high accuracy in identifying stellar types, such as:

• G-type dwarfs: Correctly identified at temperatures near 5,500K
• A-type stars: Correctly identified at ~8,500K
• Key Features: Successfully identifies the Balmer series and Fe I absorption lines

---

🔮 Future Plans

This is Version 1 of the system. In the future, we plan to:

1. Expand the training to a full 285 steps with a larger dataset.
2. Build a specialized Stellar Expert system that can identify rare chemical anomalies in distant stars.

---

Contact

Liyakhath Shaik Email: [email protected]