3D LLM Diffusion: Tri-Modal Qwen v11
This release contains the trained components required for the project's
text/XRD/crystal shared representation and Qwen-controlled Crystalite runtime.
Qwen owns both natural-language boundaries: it parses user requests into
auditable crystal conditions and produces grounded descriptions from facts
computed from the selected CIF. Crystalite generates structures; deterministic
powder diffraction physics generates XRD. The Neutral Composition Planner and
charge-balance gate are enabled for de novo generation by default. There is no
Robocrystallographer or template-prose runtime path.
Runtime Contract
1raw text -> Qwen parse + Qwen latent adapter -> v11 256-d shared hub
2 |
3 conditioned Crystalite
4 |
5 selected CIF
6 / \
7 simulated XRD facts -> Qwen text
The planner fixes a formally charge-feasible composition before Crystalite
denoises coordinates and the lattice. This is a generation constraint, not a
substitute for structure-aware validity, MLIP relaxation, or DFT validation.
End-to-End Qwen Runtime Validation
Gadi GPU job 174157319 exercised both production text routes with the pinned
Qwen revision and exited successfully:
| Route | Qwen parse | Crystal anchor | Text output |
|---|
| Retrieval | formula=NaCl, crystal_system=cubic | exact formula and crystal-system match, MP-20 mp-22862 | Qwen grounded on saved CIF facts |
| De novo | formula=NaCl, bandgap=1.1 eV, crystal_system=cubic | two charge-balanced Planner + Crystalite candidates accepted | Qwen grounded on the selected generated CIF |
The short 12-step de-novo integration sample is NaCl but has P1 symmetry, so it
does not demonstrate cubic-condition accuracy. Parsed conditions, generated
CIFs, charge audits, deterministic XRD, grounding facts, and Qwen descriptions
are preserved under metrics/qwen_anyio_smoke/. This smoke test validates the
runtime wiring; GenBench and property-control evaluations remain the quality
evidence.
The Qwen weights are not redistributed here. Use
Qwen/Qwen3.6-35B-A3B at exact
revision
995ad96eacd98c81ed38be0c5b274b04031597b0 with
transformers==5.14.1.
Released Components
| Component | Purpose |
|---|
checkpoints/trimodal_v11/trimodal_best.pt | final text/XRD/crystal shared encoder |
checkpoints/qwen_latent_adapter/qwen_latent_adapter_best.pt | pooled Qwen hidden state to final 256-d hub |
checkpoints/generator_v12/best.pt | multi-property conditioned Crystalite denoiser |
checkpoints/qwen_crystal_clip/clip_best.pt | base Qwen/crystal alignment checkpoint |
checkpoints/property_encoder/encoder_prop_best.pt | crystal property encoder |
checkpoints/qwen_property_projector/projector_step003000.pt | crystal latent to Qwen soft-token projector |
checkpoints/mace_eform/mp20_eform_run-7.model | energy guidance/evaluation model |
source/source_only_worktree_20260720.tar.gz | exact CSCS source snapshot used for migration |
source/github_70a66c2e4ce1417cfef378e4c01045d753900d0b.tar.gz | exact final Qwen runtime source at GitHub commit 70a66c2e |
manifest/ARTIFACT_MANIFEST.json | sizes and SHA-256 checksums |
Large embedding tables and training/evaluation data are released separately in
Yangfan78/3D_LLM_Diffusion-trimodal-assets-v11.
Qwen Latent Adapter Validation
The 2048-to-256 adapter was selected at epoch 99 on 9,047 held-out paired
texts.
| Metric | Value |
|---|
| Mean cosine to frozen v11 target | 0.97135 |
| Retrieval R@1 | 99.36% |
| Retrieval R@10 | 100.00% |
These metrics measure recovery of the frozen v11 text-hub targets; they are not
generated-crystal validity or property scores. The checkpoint includes the
exact Qwen revision, environment versions, hyperparameters, row counts, and
input/checkpoint SHA-256 hashes.
Tri-Modal Evaluation
The paired MP20 synthetic-XRD test split contains 9,046 examples.
| Direction | Recall@10 |
|---|
| Text to crystal | 0.9667 |
| Crystal to text | 0.9423 |
| XRD to crystal | 0.9938 |
| Crystal to XRD | 0.9992 |
| Text to XRD bridge | 0.8737 |
| XRD to text bridge | 0.8332 |
| Corrupted XRD to crystal | 0.9000 |
These are retrieval results on simulated diffraction, not measured-XRD or
generated-structure property accuracy. Generated candidates still require
independent structural, MLIP, and preferably DFT validation.
Planner + ORB-v3 GenBench
The completed all-2,500 run uses the Neutral Composition Planner, Crystalite FK
K8 sampling (lambda=4), independent ORB-v3 pre-relaxation, and the pinned
LeMat-Bulk reference revision. It is a non-hybrid generated pool.
| Metric | Result | Denominator |
|---|
| Official validity | 96.88% | 2,422 / 2,500 |
| Uniqueness | 80.14% | 1,941 / 2,422 |
| Novelty | 79.15% | 1,917 / 2,422 |
| SUN | 0.50% | 12 / 2,422 |
| SUN + MetaSUN | 26.51% | 642 / 2,422 |
| Stable after GenBench three-MLIP relaxation | 6.28% | 152 / 2,422 |
| Mean energy above hull after three-MLIP relaxation | 0.1626 eV/atom | 2,422 |
| Mean relaxation displacement RMSE | 0.1643 A | 2,422 |
The run passes the project's validity, novelty, metastability, and displacement
targets. It does not pass the uniqueness, stable-rate, or hull-energy targets.
The full CIF pools, audit trail, benchmark pickle, source snapshot, hashes, and
metric JSON are in the companion dataset under
evaluation/planner_orb_genbench_all2500_20260719/.
Source
