MODAR-data / source_data /tables /table_S19_part1.csv
Coooolin's picture
Release MODAR weights and data without source code
5ae6cc4 verified
Raw History Blame Contribute Delete
2.18 kB
Checkpoint,Encoder and parameters,Hardware and batch,Learning rate and length,Where it is used
Peak-list model,"(d, I) peak-set encoder; 385,764,352 parameters. Decoder initialized from the raster-encoder base model; peak-set encoder trained in this run. Experimental prefix perturbations enabled.",4 GPUs; per-GPU batch 12; accumulation 4; global batch 192,"Peak learning rate 1×10−4; warmup 1,000 steps; cosine decay; 20,000 steps (about 1.11 passes)","CNRS/RRUFF, the public-337 ladder and the LiGaCl4 comparison. No size router."
Raster-encoder base model,"One-dimensional convolutional spectrum encoders; 309,495,936 parameters. No experimental prefix perturbation in base training.",6 GPUs; per-GPU batch 16; accumulation 2; global batch 192,"Peak learning rate 3×10−4; warmup 2,000 steps; cosine decay; 54,000 steps (about 3.0 passes)",Full simulated ladder and aligned neutron analysis (Figs. 3–4); base-model size comparisons (Fig. 5); clean-input baseline for synthetic-artefact comparisons.
Large-cell specialist,"Raster encoder; 309,495,936 parameters. Initialized from the raster-encoder base model; trained on 362,673 cells with more than 20 sites.",6 GPUs; per-GPU batch 16; accumulation 2; global batch 192,"Peak learning rate 1×10−4; warmup 500 steps; 15,000 steps (about 7.9 passes)",Known-size comparison in Fig. 5d–f; clean-input large-cell baseline in Fig. 6.
Base noise-adapted model,"Raster encoder; 309,495,936 parameters. Initialized from the raster-encoder base model; experimental prefix perturbations enabled.",2 GPUs; per-GPU batch 16; accumulation 2; global batch 64,"Peak learning rate 5×10−5; warmup 300 steps; 6,000 steps","Synthetic-artefact ladder and base frozen-pool analyses in Fig. 6; laboratory powder series, including LiCoO2 in Fig. 7c."
Noise-adapted large-cell specialist,"Raster encoder; 309,495,936 parameters. Initialized from the large-cell specialist; experimental prefix perturbations enabled on the large-cell training subset.",2 GPUs; per-GPU batch 16; accumulation 2; global batch 64,"Peak learning rate 5×10−5; warmup 300 steps; 4,000 steps","Large-cell robustness and frozen-pool analyses in Fig. 6e,f."