btraven/scbert-panglao-pretrain

scBERT pretrained checkpoint (panglao_pretrain.pth) — unmodified mirror

A byte-identical mirror of the pretrained scBERT checkpoint from TencentAILabHealthcare/scBERT (Yang et al., scBERT as a large-scale pretrained deep language model for cell type annotation of single-cell RNA-seq data, Nature Machine Intelligence 4, 852–866 (2022), doi:10.1038/s42256-022-00534-z).

This repo exists so the file can be fetched by a script and verified by hash: upstream distributes it only through a WeChat Drive link that needs a login (or on request by email), and the only public copy is a third-party Google Drive re-upload (see Provenance). The file is not modified, retrained or converted.

file	bytes	sha256	md5
panglao_pretrain.pth	87,631,538	aa27109136580b24a1b985f63d0ba31f22490d09e71b13d1793cbff38619dfe3	1ff8352125f077de108f54d51bbcfecc

Provenance

• Downloaded on 2026-10-01 from the Google Drive file 1_Pgk_o8AtQtoXr_ZLQx0eJYoSWzjxC8f. It was re-uploaded to Google Drive by a GitHub user in scBERT issue #46 (2024-09-17), apparently after obtaining it from the author by email (the repo owner had pointed requesters to fionafyang@tencent.com). It is not verified by Tencent.
• No checksum is published upstream; the hashes above are of the bytes as downloaded.
• Cross-check: every tensor matches the independently published conversion kaichenxu/cape_scbert @ 88256813 — 90 of 91 bit-identical, and pos_emb.emb.weight identical after its float64 → float32 cast. That copy may come from the same re-upload, so this corroborates but does not prove that these are the original weights.
• Contents: a torch.save dict with epoch (5), model_state_dict (91 tensors, 8.49M parameters: 200-dim Performer, 7 expression-bin tokens, a 16,907 × 200 float64 gene2vec positional embedding), optimizer_state_dict, scheduler_state_dict, losses. Tensors were saved on cuda:0; load with torch.load(path, map_location="cpu") on a machine without a GPU.

Source code

The code that trains and uses this checkpoint is scBERT itself: TencentAILabHealthcare/scBERT@262fd4b9.

Licence

GPL-3.0 (see LICENSE, copied verbatim from scBERT@262fd4b9). Upstream's licence states:

Copyright (c) 2023 Tencent Inc. This copyright and license covers both the source code and model weights referenced for download in the README file.

Why it is mirrored

Used by omni-scfm to reproduce the scBERT arm of Ahlmann-Eltze, Huber & Anders (2025), Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselines, Nature Methods.