Move all current paper-table3 checkpoint paths to paper-table1, except paper-table3-finetune-properties-from-joint-force-conditioned and paper-table3-finetune-properties-from-joint-properties-force-conditioned, which move to paper-table10. Checkpoint contents are unchanged.
Back up paper-selected SCOPe no-force-conditioning checkpoint at step 125000
SCOPe-2k protein model finetuned from Zatom-2 (IV) generation + structure pretraining without force conditioning. Selected global step 125000 (25000 finetuning steps) in the current Zatom-2 manuscript. SHA256: 4dc8df979d7ef6aa205adf77a6d742cb304e1e22d8b7d808056f2d5296103ed6
Back up paper-selected SCOPe model IV checkpoint at step 120000
SCOPe-2k protein model finetuned from generation + structure pretraining with force conditioning. Selected global step 120000 (20000 finetuning steps) in the current Zatom-2 manuscript. SHA256: d146c9aa1c3270c6c976cbcdff0f77734c0bddfcdf1be8c53b2988699b63af00
Back up paper-selected SCOPe model V checkpoint at step 130000
SCOPe-2k protein model finetuned from generation + structure + MLIP pretraining with force conditioning. Selected global step 130000 (30000 finetuning steps) in the current Zatom-2 manuscript. SHA256: 803faeab28541bdae974d8f86cc9265bd74728c949e6b2e435504583d58e5563
Consolidate paper checkpoints under paper-ready-checkpoints
Remove 14 root-level checkpoint pointers only after verifying byte-identical nested copies by remote LFS SHA-256. Canonical paper artifacts now live under paper-ready-checkpoints/.
Paper-only, non-production checkpoint used for the manuscript's nominal epoch-240 auxiliary results. Raw embedded checkpoint epoch: 236. SHA-256: f6bf4fb1ef5e1d88c2dc90de43216163db068a48b8304406ed2bf0b15edcac44.
Non-production paper artifact for Table 2: QM9 + MP20 unconditional without force conditioning. Embedded current_epoch=9925, global_step=119112. SHA-256: 412e14210ffe4bb7d50d87c64bdea658a767238ea87f56888dd6c06734d02e55.
Paper Figure 1 ablation artifact; not a production model. Zatom-2-L trained on fixed OMol25 500k + OMat24 500k supports with joint-properties tasks and force conditioning under a 40-epoch schedule. Source epoch 39, global step 48744. SHA-256 244389d755dce61bdcd268c0fbb35be99412ac56c885721ecfd628567025f576.
Paper Figure 1 ablation artifact; not a production model. Zatom-2 trained on fixed OMol25 500k + OMat24 500k supports with joint-properties tasks and force conditioning under a 40-epoch schedule. Source epoch 39, global step 48801. SHA-256 6653853c6c8e8d33c69837bfde8e7e60dd16c412c07e1cf5540c8d6815d2f0ed.
Upload Table 3 properties fine-tune from joint checkpoint
Paper Table 3 artifact; not a production model. Zatom-2 properties fine-tuning initialized from joint force-conditioned pretraining. The checkpoint completes epoch 120: 80 pretraining epochs plus 40 fine-tuning epochs. Source epoch 119, global step 146352. SHA-256 5c12d4f631a4017ea5dacc3fe9704a34ef26c37ad46ebfa3b3a6c50f50e10bbf.
Upload Table 3 properties fine-tune from joint-properties checkpoint
Paper Table 3 artifact; not a production model. Zatom-2 properties fine-tuning initialized from joint-properties force-conditioned pretraining. The checkpoint completes epoch 120: 80 pretraining epochs plus 40 fine-tuning epochs. Source epoch 119, global step 146363. SHA-256 f2b79ed6673b54dbedc55ab7f93db6fb353c4060e9d5d800e5fdf212e50b2752.
Paper Figure 1 ablation artifact; not a production model. Zatom-2-L trained on fixed OMol25 500k + OMat24 500k supports with joint-properties tasks and force conditioning under a 20-epoch schedule. Source epoch 19, global step 24368. SHA-256 7375b21f3765904ef0938859b00c2ca960ed4f8a9de4ddc451d5ed073dbab282.
Paper Figure 1 ablation artifact; not a production model. Zatom-2 trained on fixed OMol25 500k + OMat24 500k supports with joint-properties tasks and force conditioning under a 20-epoch schedule. Source epoch 19, global step 24400. SHA-256 e78da9d1a90d67e893740cc9af8609cda16e88aa3e635da8bbe0b2a9fbd3a4de.
Paper ablation artifact; not a production model. Zatom-2-L trained on OMol25 + OMat24 joint-properties tasks with force conditioning under an 80-epoch schedule. Source step 96750, current_epoch 79. SHA-256 ebaa4afa55b7ab4d51df819ad45a95182cf03435a05460a37d1e64242c79b386.
Paper ablation artifact; not a production model. OMol25 + OMat24 joint-properties training with force conditioning, 80-epoch schedule. Source step 97500, current_epoch 79. SHA-256 7e265ebdc8190af066f2165ea2d28a8aea92525438ddcd61a7b313d9a7056963.
Paper ablation artifact only; not a production-ready model. Zatom-2-S trained for 80 epochs on the OMol25 unconditional task without force conditioning. Source step 97500; SHA-256 6f99f4cbec8a38439ca71d162c806e90950859f14130972a4ed18d05d526c3e3.
Paper ablation artifact only; not a production-ready model. Zatom-2-S trained for 80 epochs on the OMol25 unconditional task with force conditioning. Source step 97500; SHA-256 895d67fcc784fd14aadb00c7ad507ef4c37ce4b25ea1927c11521acbfacdb0d4.
These two files are paper ablation checkpoints, not production-ready Zatom-2 releases. They remain available under explicit Table 1 and Table 3 ablation filenames.
Paper ablation artifact only; not a production-ready model. Zatom-2 trained for 80 epochs jointly on OMol25 and OMat24 unconditional and structure-prediction tasks with force conditioning. Source step 97500; SHA-256 a47edf3608c30bcb4c9273b7a71bc81b0ef7a282d6c0aa4a741463176c8590c8.
Paper ablation artifact only; not a production-ready model. Zatom-2-S trained for 80 epochs on OMol25 unconditional and structure-prediction tasks with force conditioning. Source step 104250; SHA-256 7a08a2df01b1d1a2176586a231aa0e0a19c707ee1c6be673ea35d2dda4a08ab5.
Joint OMol25 and OMat24 unconditional and structure-prediction pretraining with force conditioning. Training completed epoch 80; source step 97500; SHA-256 a47edf3608c30bcb4c9273b7a71bc81b0ef7a282d6c0aa4a741463176c8590c8.
OMol25 unconditional and structure-prediction pretraining with force conditioning. Training completed epoch 80; source step 104250; SHA-256 7a08a2df01b1d1a2176586a231aa0e0a19c707ee1c6be673ea35d2dda4a08ab5.
