raman-01-1.7B

Raman-01: Pocket Physics Solver LLM

Raman-01 is a compact, RL-finetuned LLM tailored specifically for solving physics problems. Built to be lightweight and easily deployable, it excels at tackling easy-to-medium difficulty physics questions across diverse domains.

🚀 Model Overview

• Base Model: Qwen3-1.7B (cold-start supervised fine-tuning)

  • SFT Data: 1500 samples covering kinematics, electromagnetism, acoustics, and other fundamental physics domains.
  • SFT Training: 3 epochs, achieving a loss reduction to ~0.3.

• Reinforcement Learning (GRPO) Fine-tuning:

  • Dataset: Single carefully-selected medium-difficulty physics sample. (1-shot RLVR)
  • Training Steps: 70 GRPO steps
  • Reward Progression: Improved from an initial reward of 0.1 to 0.8 by training completion.
  • Methodology closely follows the training style pioneered by the DeepSeek-R1 model but on a smaller, targeted scale.

🎯 Use Cases

Ideal for quick, reliable physics problem-solving in:

• Mobile or edge deployments requiring minimal computational resources.
• Educational tools and pocket-sized physics assistants.
• Rapid prototyping of physics-related AI applications.

📊 Performance & Benchmarks

• Demonstrates strong performance on easy-to-medium difficulty physics problems, especially effective in fundamental physics domains such as mechanics, electromagnetism, and basic acoustics.
• Evaluations recommended using PhyBench or similar physics-focused benchmarks for accuracy assessment.

⚙️ Deployment

• Compact model size (1.7B parameters) optimized for rapid inference on CPU/GPU.
• Easily deployable via popular frameworks (PyTorch, Hugging Face, vLLM).

📝 Limitations

• Specialized on easy-medium physics problems; performance on complex, multi-domain or advanced theoretical questions may vary.
• Trained primarily to demonstrate capabilities in a constrained setting; broader fine-tuning advised for production-grade deployments.

🔗 Citation

If you utilize Raman-01, please acknowledge accordingly:

@misc{raman01,
  author = {Sai Praneeth Diddigam},
  title = {Raman-01: Compact RL-Enhanced Physics Solver},
  year = {2025},
  note = {RL (GRPO) finetuned on Qwen3-1.7B}
}

---

Developed to be your go-to physics solver, Raman-01 packs powerful performance into a conveniently small footprint.