matsci-llm-anon/llm-materials-code

How Large Language Models Learn Materials Science — code

Notebooks for fine-tuning and evaluating large language models on four materials science tasks. Released for double-blind review; author and affiliation information is withheld.

Tasks

Task	Type
Bandgap prediction	regression (eV)
Dielectric constant prediction	regression
Crystal system classification	7-class classification
MatKG link prediction	knowledge-graph completion

Notebooks

Notebook	Purpose
01_Predibase_Finetuning_Inference.ipynb	LoRA fine-tuning and inference for the open-weights models; contains the task prompt templates
02_GPT_Finetuning_Inference.ipynb	OpenAI fine-tuning and inference, including the longitudinal GPT runs
03_Embedding_Extraction.ipynb	Layer-wise embedding extraction and the supervised probes
04_Results_Analysis.ipynb	Metric computation, entropy and aggregation
05_Figures.ipynb	Figure generation

Prompt templates for all four tasks are defined inline in notebooks 01 and 02.

Datasets

Dataset	Task
matsci-llm-anon/materials-bandgap-prediction	regression
matsci-llm-anon/materials-dielectric-prediction	regression
matsci-llm-anon/materials-crystal-system-classification	classification
matsci-llm-anon/matkg-link-prediction	link prediction

Fine-tuned adapters

19 LoRA adapters across Llama-2-7B, Llama-3-8B, Mistral-7B and Mixtral, one per task plus a combined multi-task variant, at matsci-llm-anon/<Base>-ft-<Task> — for example matsci-llm-anon/Llama2-ft-BandGap.

Environment

pip install -r requirements.txt

Note on scope

Per-run inference outputs and the full result CSVs are not mirrored here; they are archived with the journal version of this work.

Citation

@inproceedings{anonymous2026learn,
  title={How Large Language Models Learn Materials Science},
  author={Anonymous},
  booktitle={Under review},
  year={2026}
}