Speaker
Description
The application of large language models (LLMs) in scientific domains, particularly materials discovery, remains underexplored compared to general natural language tasks. This study presents a systematic fine-tuning of the DeepSeek-R1 model using Unsloth’s memory-efficient framework and domain-specific supervised instruction techniques. Three publicly available datasets—GPT-Narratives-for-Materials, MDF-LLM Resources, and MatChem-LLM—were curated and harmonized to develop a fine-grained instruction-response corpus relevant to material science properties, experimental conditions, and compositional analysis.
The fine-tuning process leveraged SFTTrainer with instruction tuning templates, a sequence length of 2048 tokens, and efficient gradient accumulation to adapt DeepSeek-R1 for expert-level retrieval, summarization, and hypothesis generation tasks in material science. Preliminary evaluations using a simulated GLUE-style protocol, including BLEU-based output quality assessment, indicate significant improvement in domain-specific language generation, with BLEU scores exceeding expectations on validation samples.
This work demonstrates that strategically fine-tuned LLMs can serve as foundational models for AI-driven material discovery knowledge bases, potentially accelerating hypothesis testing, experimental planning, and scientific reasoning. Future directions include integrating regression-based materials property prediction benchmarks and open-domain QA fine-tuning for broader applicability.
