HotPOT checkpoints (anonymous submission)

Weights for the 24 HotPOT entries of Table 1: two pre-trained models (E1000) and twelve fine-tuned models (E200). Each file stores only the model tensors and loads with hotpot.load_checkpoint from the anonymous code repository (https://anonymous.4open.science/r/HotPOT-BA73/). checkpoints.json maps every Table 1 entry to its file, test rows, rollout length, expected error and evaluation command; the same table is in docs/CHECKPOINTS.md of the code repository.

File Scale Stage Bytes SHA-256
HotPOT-T_pretrain_E1000.pt 17M pretrain E1000 391,847,761 4c600e1f7b0314bea980c6daa839ce66ad100f57467c5b22cdcab6d46f8cbc8d
HotPOT-S_pretrain_E1000.pt 90M pretrain E1000 2,074,225,745 3800e428ef6d75798bf9fb06a6b9eb591d2f7e5eac27c00d9a733d0b78acc62a
HotPOT-T_ns2d_fno_1e-5_finetune_E200.pt 17M finetune E200 391,855,807 8280ad5c502a73d7cd68cf9e7f59692890eb84ee3988ccc20a3c7d02eba4fd70
HotPOT-S_ns2d_fno_1e-5_finetune_E200.pt 90M finetune E200 2,074,236,287 2701d6e1bda606739ff413cd7e60fffcb42f395c6e8478e0ba60dafde9e55447
HotPOT-T_ns2d_fno_1e-3_finetune_E200.pt 17M finetune E200 391,855,807 8aef17d136b6166a9ce0b22d561deb54c448431a70fd7d7286a4de6e85719dfb
HotPOT-S_ns2d_fno_1e-3_finetune_E200.pt 90M finetune E200 2,074,236,287 35165448e148261031df322e19e8a5c62aed7bcebf0d5b9fa02bdf634eb4bbe5
HotPOT-T_ns2d_pdb_M1e-1_eta1e-2_zeta1e-2_finetune_E200.pt 17M finetune E200 391,867,115 046e2b3e65a362f97bf5cfb9cc5fee36681a81c1f6f4411c33da8990231ce9b6
HotPOT-S_ns2d_pdb_M1e-1_eta1e-2_zeta1e-2_finetune_E200.pt 90M finetune E200 2,074,250,987 88d5bfa1cd6c91751916afbf8a05dfb0a05023be5954862637f149c3807880e1
HotPOT-T_swe_pdb_finetune_E200.pt 17M finetune E200 391,852,123 4f62297100694ee37223a55336978591fb33c10fed094b6f270bc1a505b71d31
HotPOT-S_swe_pdb_finetune_E200.pt 90M finetune E200 2,074,231,387 b37606009586b45494cf74b424fff37c716d864d2c7531723571c0ef56b9fc27
HotPOT-T_dr_pdb_finetune_E200.pt 17M finetune E200 391,851,445 e319ed247c166a80854c9aa78479f48e14328fd6dd53cf9acf93e9ba7aa35776
HotPOT-S_dr_pdb_finetune_E200.pt 90M finetune E200 2,074,230,581 9feb8ade90f5d170e790be2837deb5ccb289ad7b3d93264cb8cc370f8a7842d9
HotPOT-T_cfdbench_finetune_E200.pt 17M finetune E200 391,852,737 e164bc3e6d52ef251545dfac8ead806a6725c75632c7c26e739a424affd5d06e
HotPOT-S_cfdbench_finetune_E200.pt 90M finetune E200 2,074,232,193 6388299d9eaef0e210d6546e406e80d1990121f195b467b1a5b772b5305f2c85
from hotpot import build_model, load_checkpoint
model = build_model(scale=17)
load_checkpoint(model, "HotPOT-T_pretrain_E1000.pt")

Single-expert files

Fine-tuning updates the shared history encoder and the expert of the target family only; the experts of the other families stay frozen and are never executed for that family. A deployed single-task model therefore needs the shared encoder and one expert. single_expert/ holds such a copy of each of the twelve fine-tuned models, written by tools/prune_checkpoint.py and loaded by the same hotpot.load_checkpoint. The kept tensors are bitwise identical to those in the full files, and predictions agree bit for bit with the full files in every case we ran (the five non-CFDB settings at the 17M scale and SWE at the 90M scale, on a fixed input).

Model Setting Expert kept Stored elements File bytes Full file bytes Reduction
HotPOT-T NS 1e-5 fno 16,776,785 67,160,183 391,855,807 82.9%
HotPOT-T NS 1e-3 fno 16,776,785 67,160,183 391,855,807 82.9%
HotPOT-T CNS (0.1, 0.01) cns 17,178,004 68,767,259 391,867,115 82.5%
HotPOT-T SWE swe 16,770,577 67,133,755 391,852,123 82.9%
HotPOT-T DR dr 16,903,188 67,663,911 391,851,445 82.7%
HotPOT-T CFDB cfdb 38,326,966 153,382,427 391,852,737 60.9%
HotPOT-S NS 1e-5 fno 89,590,881 358,434,071 2,074,236,287 82.7%
HotPOT-S NS 1e-3 fno 89,590,881 358,434,071 2,074,236,287 82.7%
HotPOT-S CNS (0.1, 0.01) cns 91,185,956 364,817,435 2,074,250,987 82.4%
HotPOT-S SWE swe 89,584,673 358,407,355 2,074,231,387 82.7%
HotPOT-S DR dr 90,112,036 360,516,407 2,074,230,581 82.6%
HotPOT-S CFDB cfdb 205,620,838 822,581,403 2,074,232,193 60.3%

The CFDB file keeps all three geometry paths (cavity, cylinder, tube). For reference, the MoE-POT-T and MoE-POT-S models store 30M and 166M parameters (about 121 MB and 664 MB in FP32); their routing is input-dependent, so all experts must be kept for a single task.

SHA-256 of the single-expert files:

File SHA-256
single_expert/HotPOT-T_ns2d_fno_1e-5_finetune_E200.single.pt 916e5fb4f91346dc109e1b4086c5763e21580e8ef99b1d307ce32737f05c604f
single_expert/HotPOT-T_ns2d_fno_1e-3_finetune_E200.single.pt b9cfef12e2bb0c043a91ad95a3bfb71fb93de821598e77f940dc4e084f50ea17
single_expert/HotPOT-T_ns2d_pdb_M1e-1_eta1e-2_zeta1e-2_finetune_E200.single.pt 70ba5036a1af8a717dfae0d621595b2699cd1e6a34980eed9aa7209be1d24ead
single_expert/HotPOT-T_swe_pdb_finetune_E200.single.pt cca4f8545d11093ac8598208609a5443b757ef1bd5b380345fc486ee51d90bdf
single_expert/HotPOT-T_dr_pdb_finetune_E200.single.pt 3147b86eea523ad0afe2f20300e022f21193e59cbef3e00187f083d802c3f094
single_expert/HotPOT-T_cfdbench_finetune_E200.single.pt 8f6ab565eecfb3bfba977aeee8d750dee0b17c83a9c1457f6d98c5356095d259
single_expert/HotPOT-S_ns2d_fno_1e-5_finetune_E200.single.pt 80fb1d00a038a2222a061692dd6708f974aaa7b70b17a7506be54b197eb2e432
single_expert/HotPOT-S_ns2d_fno_1e-3_finetune_E200.single.pt a4dead5a3bfab4bb77db56184617d70e2359215601d7938eebb74219d1dfbb0d
single_expert/HotPOT-S_ns2d_pdb_M1e-1_eta1e-2_zeta1e-2_finetune_E200.single.pt dac7eec870ecca4a12c777e3b09b940001de65efb790f9d9c8d3e1dcc9d964a4
single_expert/HotPOT-S_swe_pdb_finetune_E200.single.pt 6ecc27ab4b009ec6c59bf1a15597e34514a08bd4e2875890b9ac20179e654e6b
single_expert/HotPOT-S_dr_pdb_finetune_E200.single.pt 63401ddec179f6c2cb218d14e15ee504e08e1b5ea4fb3248b4b6c1098010f2c5
single_expert/HotPOT-S_cfdbench_finetune_E200.single.pt d1b096b113a540afbc30c9a7461ba72f55921b12ce9c887e61f0af11d7de76be
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support