The dataset could not be loaded because the splits use different data file formats, which is not supported. Read more about the splits configuration. Click for more details.
Error code: FileFormatMismatchBetweenSplitsError
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
3D LLM Diffusion: Tri-Modal Assets v11
This dataset repository contains the large data artifacts required to reproduce
the v11 text/XRD/crystal routing layer and train the released Qwen latent
adapter. It accompanies
Yangfan78/3D_LLM_Diffusion-trimodal-qwen-v11.
Contents
| Path | Contents |
|---|---|
embeddings/trimodal_cond_emb_v11/ |
final v11 train/validation/benchmark latent tables |
corpus/mm_corpus/ |
27,136 train and 9,047 validation text records plus alignment targets |
data/mp20_props/ |
processed MP20 property targets used by the multi-property model |
data/mp20_raw/ |
portable train/validation/test CSV splits used by this project |
prompts/prompt_bench_v1/ |
fixed conditional-generation prompt benchmark |
xrd/xrd_hard_negatives/ |
hard-negative metadata/artifacts |
xrd/xrd_cache_cscs_20260720.tar |
45,229 cached synthetic XRD tensors, packed to preserve HPC inode quota |
evaluation/completeness_v11_20260702/ |
v11 completeness evaluation artifacts |
evaluation/planner_orb_genbench_all2500_20260719/ |
complete 2,500 raw/ORB CIF pools, generation and relaxation audits, pinned LeMat GenBench results, and intermediate benchmark artifacts |
manifest/ARTIFACT_MANIFEST.json |
sizes and SHA-256 checksums |
The XRD cache contains simulated powder patterns. It must not be interpreted as an experimental diffraction dataset. Downstream users are responsible for checking the provenance and usage terms of upstream Materials Project data.
Qwen Adapter Pairing
Each corpus split has complete one-to-one ID coverage with its 256-dimensional target table:
| Split | Text rows | Target rows | Shared IDs |
|---|---|---|---|
| Train | 27,136 | 27,136 | 27,136 |
| Validation | 9,047 | 9,047 | 9,047 |
The input-side adapter uses pooled hidden states from
Qwen/Qwen3.6-35B-A3B revision
995ad96eacd98c81ed38be0c5b274b04031597b0 and maps them into the final v11
256-dimensional shared hub.
The released epoch-99 adapter reaches mean target cosine 0.97135, R@1
0.99359, and R@10 1.00000 on the 9,047 validation pairs. These are
latent-target retrieval metrics, not crystal generation metrics.
Full Planner + ORB-v3 Evaluation
The all-2,500 evaluation completed with 96.88% official validity and 0.1643 A
mean GenBench relaxation displacement RMSE. Strict SUN is 0.50%; uniqueness is
80.14%; the post-three-MLIP stable rate is 6.28%. These lower metrics are
retained explicitly rather than collapsed into the validity result. See the
evaluation directory for exact denominators, source/pool hashes, target checks,
and the pinned LeMaterial/LeMat-Bulk revision.
Source
- Downloads last month
- 857