alm-edit/README.md

metadata

license: apache-2.0
tags:
  - materials
  - diffusion
  - crystal-structure-prediction
  - editing
  - mattergen

ALM Edit · crystal-structure prediction + text-conditioned editing

ALM Edit conditions a from-scratch CSP-mode MatterGen decoder (csp_backbone) on an input structure + text instruction via a producer-consumer bridge: K=8 [atoms_i] soft tokens plus prompt context feed a learnable-query producer (M=16), whose tokens the cross-attention consumer reads inside the decoder. It performs crystal-structure prediction and the ALM Bench editing tasks.

This is a full-finetuning run: it bundles a fully fine-tuned Qwen3-8B (llm_full_ft/qwen3_state_dict.pt, ~16 GB), which the loader auto-detects and uses in place of any LoRA. Contents: atoms_mapper.pt (producer-consumer bridge) + llm_full_ft/

• projector_and_state.pt + the csp_backbone/ decoder dir (config + last.ckpt).

Crystal-structure prediction: match rate MR (%, ↑) and RMSE (Å, ↓) to the MP-20 and MPTS-52 test sets at K=1 and best-of-K=20 (95% CIs in the paper):

Model	MP-20 MR@1	RMSE@1	MR@20	RMSE@20	MPTS-52 MR@1	RMSE@1	MR@20	RMSE@20
CDVAE	33.90	0.1045	66.95	0.1026	5.34	0.2106	20.79	0.2085
DiffCSP	51.49	0.0631	77.93	0.0492	12.19	0.1786	34.02	0.1749
FlowMM	61.39	0.0566	n/a	n/a	17.54	0.1726	n/a	n/a
CrystaLLM-large	58.70	0.0408	73.97	0.0349	19.21	0.1110	33.75	0.1059
CrystalFlow	62.02	0.0710	78.34	0.0577	21.00	0.1613	37.81	0.1584
OMatG	63.75	0.0720	n/a	n/a	25.15	0.1931	n/a	n/a
MCFlow-L	64.08	0.0561	76.08	0.0383	27.16	0.1401	41.45	0.1296
ALM Edit	45.6	0.021	83.2	0.034	22.7	0.022	45.7	0.038
ALM Gen + T2C-FK	22.3	0.025	41.0	0.012	6.0	0.040	10.0	0.011

ALM Edit sets SoTA RMSE and best-of-K=20 match rate on both benchmarks (it learns a valid polymorph distribution); it sees composition + space group at train time, only composition at inference.

ALM Bench directional editing per property (E_f, ρ, V) and direction (↑/↓), plus polymorph/doping/strain (N=7×1000; every metric scores invalid generations, trivial lattice rescalings, and unphysical relabelings as failures). Frontier LLMs were prompted to read/write CIFs:

Model	E_f↑	E_f↓	ρ↑	ρ↓	V↑	V↓	Polymorph	Doping	Strain
ALM Edit	0.613	0.624	0.353	0.367	0.451	0.355	0.224	0.879	0.151
GPT-4o	0.505	0.469	0.024	0.127	0.081	0.018	0.040	0.007	0.000
GPT-4.1	0.465	0.496	0.007	0.239	0.276	0.040	0.083	0.003	0.000
GPT-5.2	0.437	0.414	0.058	0.244	0.006	0.032	0.118	0.002	0.000

Text-conditioned generation (Application = LLM-judged fit; Describe/OOD = composition & structure consistency, N=7×1000):

Model	Application	Describe (Comp.)	Describe (Struct.)	OOD (Comp.)	OOD (Struct.)
ALM Edit	0.423	0.730	0.412	0.474	0.231
GPT-4o	0.131	0.279	0.121	0.130	0.025
GPT-4.1	0.224	0.254	0.090	0.168	0.035
GPT-5.2	0.252	0.356	0.162	0.263	0.075

Edit / generate a structure (inference):

# composition/description -> structure, using the CSP-mode backbone
alm-generate generate --alm_checkpoint alm-edit \
    --atoms_mapper alm-edit/atoms_mapper.pt --mattergen_model_path alm-edit/csp_backbone \
    --prompt "An orthorhombic perovskite of calcium and titanium." --num_samples 8 --out_dir gen_out

Evaluate (CSP M@20 / RMSE, and the ALM Bench editing tasks):

alm-eval-csp --ckpt_dir alm-edit/csp_backbone \
    --guidance_factor 0.5 --out_dir out/csp
alm-eval-almbench --alm_checkpoint alm-edit \
    --atoms_mapper alm-edit/atoms_mapper.pt --bridge_lora_dir none \
    --mattergen_model_path alm-edit/csp_backbone --guidance_factor 0.5

Links

Paper: arXiv · HuggingFace · Code: GitHub

License

Apache-2.0.

Citation

@article{edamadaka2026atomistic,
  title   = {Atomistic Language Models Understand and Generate Materials},
  author  = {Edamadaka, Sathya and Ramesh, Krithik and Li, Ju and G\'omez-Bombarelli, Rafael},
  journal = {arXiv preprint arXiv:2606.21395},
  year    = {2026}
}