DriveDNA-models: trained checkpoints behind the DriveDNA benchmark

These are the 47 PyTorch checkpoints that produced the tables of DriveDNA: A Large-Scale Multimodal Naturalistic Driving Dataset and Benchmark for Driving Style Identification (arXiv:2607.23822). Code, evaluation harness and result files: github.com/WangYuHang-cmd/DriveDNA (tag v1.1-kdd2027). Data: HenryYHW/DriveDNA.

Every file is a plain torch.save dictionary with the model state dict (model), the per-channel input normalisation statistics (mu, sd, 17 channels), and for some models an arch string. No data, identifiers or paths are stored inside. Total size 162 MB; sha256 of every file in CHECKSUMS.sha256.

Loading

import torch, sys
sys.path.insert(0, "code/model"); sys.path.insert(0, "code/eval")   # from the GitHub repository
from s1_supcon import Encoder
ck = torch.load("s1_supcon.pt", map_location="cpu", weights_only=False)
enc = Encoder(c_in=17).eval(); enc.load_state_dict(ck["model"])
x = (windows_x - ck["mu"]) / ck["sd"]      # windows_x: [N, 600, 17] float32 (features/windows_x.npy in the dataset)

To evaluate with the paper's protocol, download the dataset, run scripts/prepare_release_layout.py --models <this folder> in the GitHub repository and use the commands in its README (for example python code/eval/t4_eval.py --ckpt experiments/checkpoints/s1_supcon_s10.pt --tag _s10).

Checkpoints

Seeds: default is the seed used for the headline row; 10 / 11 are the two additional training seeds of the three-seed protocol; fe* checkpoints were trained for the fixed-evaluation-split control (App. I).

File	MB	Training command	Paper	Seed	Status	sha256
m2_mcpp6_dinov2.pt	1.82	code/model/m2_mcpp6.py --video dinov2	Sec. 6.5 / App. K (MCPP modality ablation)	default	superseded by m2_mcpp6_dinov2_v2.pt	01bc2555a33f…
m2_mcpp6_dinov2_v2.pt	1.82	code/model/m2_mcpp6.py --video dinov2 --tag _v2	Sec. 6.5 / App. K (MCPP modality ablation)	default	used	77e4276c8610…
m2_mcpp6_dinov3.pt	1.82	code/model/m2_mcpp6.py --video dinov3	Sec. 6.5 / App. K (MCPP modality ablation)	default	used	84b49188ed58…
m2_mcpp6_vjepa2.pt	1.95	code/model/m2_mcpp6.py --video vjepa2	Sec. 6.5 / App. K (MCPP modality ablation)	default	used	ede08e4f74b4…
m2_mcpp6_vjepa2_s10.pt	1.95	code/model/m2_mcpp6.py --video vjepa2 --seed 10	Sec. 6.5 / App. K (MCPP modality ablation)	10	used	36b101496423…
m2_mcpp6_vjepa2_s11.pt	1.95	code/model/m2_mcpp6.py --video vjepa2 --seed 11	Sec. 6.5 / App. K (MCPP modality ablation)	11	used	1e03cee9be3e…
m3_t5_video.pt	0.58	code/model/m3_t5_video.py (earlier version)	App. L (event forecasting with video)	default	superseded by the --video variants	867780951371…
m3_t5_video_dinov2.pt	0.58	code/model/m3_t5_video.py --video dinov2	App. L (event forecasting with video)	default	used	9b612b5d1fce…
m3_t5_video_dinov3.pt	0.58	code/model/m3_t5_video.py --video dinov3	App. L (event forecasting with video)	default	used	ed7024234b3f…
m3_t5_video_siglip2.pt	0.58	code/model/m3_t5_video.py --video siglip2	App. L (event forecasting with video)	default	used	c2ff2bff0f42…
m3_t5_video_vjepa2.pt	0.68	code/model/m3_t5_video.py --video vjepa2	App. L (event forecasting with video)	default	used	28139de138e8…
m4_clip.pt	8.85	code/model/m4_clip_align.py	Table 3 (CLIP-aligned CAN)	default	used	4dae862689a2…
m6_transfuser_lite.pt	4.57	code/model/m6_transfuser_lite.py	Sec. 6.5 / App. K	default	used	cbb148b1bab8…
m7_vidpool.pt	1.06	code/model/m7_video_probe.py	Sec. 6.5 (video-only probe)	default	used	fe8b5c4e4f1a…
s1_dann.pt	7.79	code/model/s1_dann.py	Sec. 6.6 (adversarial vehicle invariance)	default	used	06a42de118b9…
s1_supcon.pt	7.79	code/model/s1_supcon.py	Table 3 (PatchTST + SupCon); Table 4 via t4_eval.py	default	used	03d9e2215e6d…
s1_supcon_s10.pt	7.79	code/model/s1_supcon.py --seed 10	Table 3 (PatchTST + SupCon); Table 4 via t4_eval.py	10	used	b3fcc3e32478…
s1_supcon_s11.pt	7.79	code/model/s1_supcon.py --seed 11	Table 3 (PatchTST + SupCon); Table 4 via t4_eval.py	11	used	0d609eaa04a8…
s3_pred.pt	2.89	code/model/s3_personalized.py	Sec. 6.4 (FiLM-conditioned GRU)	default	used	440fac342405…
s3_pred_fe10.pt	1.09	code/model/s3_personalized.py --seed 10	Sec. 6.4 / App. H, I	10 (fixed evaluation split, App. I)	used	03cf4eb2a854…
s3_pred_fe11.pt	1.09	code/model/s3_personalized.py --seed 11	Sec. 6.4 / App. H, I	11 (fixed evaluation split, App. I)	used	f7debc6e54bc…
s3_pred_s10.pt	1.09	code/model/s3_personalized.py --seed 10	Sec. 6.4 / App. H, I	10	used	7042e2a1a5db…
s3_pred_s11.pt	1.09	code/model/s3_personalized.py --seed 11	Sec. 6.4 / App. H, I	11	used	6bb7516327e2…
s3_pred_s1z.pt	0.99	code/model/s3_personalized.py --use_s1	Sec. 6.4 (re-ID embedding as conditioning signal)	default	used	a66eb3e0c01e…
s3_pred_tf.pt	2.89	code/model/s3_personalized.py --arch transformer	App. M (Transformer predictor)	default	used	7379b7f23360…
s4_mcpp.pt	1.77	code/model/s4_mcpp.py	Sec. 6.5 (MCPP multimodal predictor)	default	used	6d5e05ede878…
s5_mdn.pt	2.02	code/model/s5_mdn.py	Sec. 6.4 (MDN distributional head)	default	used	b74ebe689f1b…
s5_mdn_fe10.pt	2.02	code/model/s5_mdn.py --seed 10	Sec. 6.4 (MDN distributional head)	10 (fixed evaluation split, App. I)	used	3002cc8938a4…
s5_mdn_fe11.pt	2.02	code/model/s5_mdn.py --seed 11	Sec. 6.4 (MDN distributional head)	11 (fixed evaluation split, App. I)	used	211d8b28b748…
s5_mdn_fe9.pt	2.02	code/model/s5_mdn.py --seed 9	Sec. 6.4 (MDN distributional head)	9 (fixed evaluation split, App. I)	used	56acc7bcdca7…
s5_mdn_s10.pt	2.02	code/model/s5_mdn.py --seed 10	Sec. 6.4 (MDN distributional head)	10	used	042727eca1de…
s5_mdn_s11.pt	2.02	code/model/s5_mdn.py --seed 11	Sec. 6.4 (MDN distributional head)	11	used	b2f84122f3e3…
w1_arcface.pt	7.79	code/model/w1_backbones.py --arch arcface	Table 3	default	used	42a93a5a2f65…
w1_arcface_s10.pt	7.79	code/model/w1_backbones.py --arch arcface --seed 10	Table 3	10	used	fdfb4cd0b963…
w1_arcface_s11.pt	7.79	code/model/w1_backbones.py --arch arcface --seed 11	Table 3	11	used	ec272e2e0d0a…
w1_ci.pt	2.66	code/model/w1_backbones.py --arch ci	Table 3	default	used	b23b10be6c45…
w1_ci_s10.pt	2.66	code/model/w1_backbones.py --arch ci --seed 10	Table 3	10	used	d4d642c8f031…
w1_ci_s11.pt	2.66	code/model/w1_backbones.py --arch ci --seed 11	Table 3	11	used	9901325c3889…
w1_itr.pt	8.08	code/model/w1_backbones.py --arch itr	Table 3	default	used	d138da649e50…
w1_itr_s10.pt	8.08	code/model/w1_backbones.py --arch itr --seed 10	Table 3	10	used	ded6da7f38e2…
w1_itr_s11.pt	8.08	code/model/w1_backbones.py --arch itr --seed 11	Table 3	11	used	a7af0a20554a…
w2_jepa.pt	7.79	code/model/w2_ssl.py --mode jepa	Table 3 (self-supervised rows)	default	used	b176373bcad9…
w2_masked.pt	7.79	code/model/w2_ssl.py --mode masked	Table 3 (self-supervised rows)	default	used	f85d6d1fb991…
w5_cvae.pt	1.43	code/model/w5_cvae.py	Sec. 6.4 / App. H (CVAE distributional head)	default	used	024a587db2d8…
w5_cvae_fe10.pt	1.43	code/model/w5_cvae.py --seed 10	Sec. 6.4 / App. H (CVAE distributional head)	10 (fixed evaluation split, App. I)	used	a55bfe000ddf…
w5_cvae_fe11.pt	1.43	code/model/w5_cvae.py --seed 11	Sec. 6.4 / App. H (CVAE distributional head)	11 (fixed evaluation split, App. I)	used	03254a56dc8f…
w5_cvae_fe9.pt	1.43	code/model/w5_cvae.py --seed 9	Sec. 6.4 / App. H (CVAE distributional head)	9 (fixed evaluation split, App. I)	unused (not referenced by any result)	80a1cec2cf2b…

License and permitted use

Released under the DriveDNA research license: non-commercial academic and research use only. These models were trained on driver identities and must not be used to identify or re-identify individual drivers, nor for insurance, employment, credit or law-enforcement scoring of individuals. Publications using them must cite the paper.

@article{drivedna2026,
  title   = {DriveDNA: A Large-Scale Multimodal Naturalistic Driving Dataset and
             Benchmark for Driving Style Identification},
  author  = {Wang, Yuhang and Li, Lingyao and Zhou, Hao},
  journal = {arXiv preprint arXiv:2607.23822},
  year    = {2026}
}