3D_LLM_Diffusion-fk-k8-lam4

3D LLM Diffusion fk_k8_lam4

This repository is the reproducible release for the fk_k8_lam4 raw LeMat-GenBench run reported in the project GitHub repository. It releases the actual generator checkpoint, the MACE formation-energy guidance model, the FK sampling configuration, all 2,500 generated CIFs, and the GenBench summary.

fk_k8_lam4 is a sampling run name, not a separate neural checkpoint. It uses v12_multiprop/best.pt with force-kernel (FK) steering: 8 particles, lambda 4.0, sigma range [0.03, 0.20], and 4 resampling checkpoints.

Released Artifacts

Direct LeMat-GenBench Result

The run uses 2,500 fixed prompts, 100 EDM steps, CFG scale 2.0, s_churn=2.0, and seed 11. CIFs were submitted without pre-relaxation and evaluated with the LeMat ORB, MACE, and UMA relaxation ensemble.

This direct raw-generator run passes 1 of the 7 project thresholds (novelty). Relaxation displacement RMSE is LeMat's per-index Cartesian displacement from the submitted structure to its MLIP-relaxed structure. It is not an RMSD to a reference crystal structure.

Reproduction Scope

reproducibility/run_fk_generation.sh regenerates the FK sampling protocol after an MP20-compatible training CSV is supplied. The current sampler derives empirical atom-count and allowed-element priors from that CSV. The original MP20 train/validation split and training text embeddings are not redistributed.

The full evaluated CIF pool is included in evaluation_cifs.tar.gz, so the published GenBench submission can be audited without rerunning generation. The source snapshot, environment lock, artifact hashes, and provenance are provided for the inference-side protocol. Retraining v12_multiprop is outside this release scope.

Source And License

The main implementation and project documentation are at Richardyangfan78/3D_LLM_Diffusion. The code and released artifacts are provided under the MIT license in this repository. Third-party code and dependencies retain their notices in NOTICE.

Limitations

This is a research candidate-generation run. The direct benchmark result is not DFT validation, does not establish synthesizability, and should not be confused with the separately curated MLIP-relaxed hybrid delivery result.