Downloads · 30 days
9
29% of all-time downloads
ZiyanZhuang/shorkie-lm-165-method-rebuild-v1.1
shorkie-lm-165-method-rebuild-v1.1 is a machine learning model from ZiyanZhuang. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for shorkie_torch. The card lists the license as apache-2.0.
Unofficial community PyTorch reproduction of ShorkieLM. Not affiliated with or endorsed by Calico or the original Shorkie authors.
Downloads · 30 days
9
29% of all-time downloads
All-time downloads
31
Public
Parameters
13.7M
54.7 MB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors54.7 MB · 100%
From the Hugging Face model README
Unofficial community PyTorch reproduction of Shorkie_LM. Not affiliated with or endorsed by Calico or the original Shorkie authors.
This repository contains the sanitized v0.1.0-rc1 safetensors release for the
v1.1 routed-L2 architecture trained with the D continuation recipe (Adam state
reset plus cosine learning-rate decay from 3e-5 to 3e-6). The label v1.1 is
specific to this reproduction and is not an official Shorkie version name.
Please cite and visit the original Shorkie paper, official repository, and official documentation first. PyTorch code and tests are maintained at ZiyanZhuang/shorkie-pytorch.
The model is not validated for clinical, diagnostic, biosafety, or other safety-critical use. It is not the supervised 5,215-track Shorkie model used for paper Figures 3-7.
N is represented by four zero DNA channels.The model was trained on an independently reconstructed 165-genome fungal corpus derived from Ensembl Fungi release 59. Repeat masking, windowing, and filtering followed the published method contract, but this corpus is not the authors' original corpus and is not included here.
The official project now lists the requester-pays 165_Saccharomycetales
corpus. Prefer the author-distributed corpus when appropriate GCP access is
available; method-rebuild is an independent audit and fallback.
On a frozen 536-window method-rebuild R64 validation plan, this checkpoint reached overall weighted PPL 3.621104. The official Shorkie_LM oracle reached 3.604430 on the same reconstructed samples and exact mask/RC plan. This model is 0.4626% higher (worse); the paired 64-kb genomic-block bootstrap PPL-ratio 95% interval is 1.004244-1.005013.
This is not reproduced supervised Shorkie performance and is not the authors'
original-corpus headline result. See evaluation.json for the machine-readable
contract and evidence boundary.
python -m pip install \
"git+https://github.com/ZiyanZhuang/[email protected]"
from shorkie_torch import load_pretrained
model = load_pretrained("/path/to/downloaded/model-directory")
The public loader verifies the SHA-256 recorded in config.json and performs a
strict safetensors load. No remote code is required. Do not load untrusted
pickle checkpoints.
8f9f7df5d0be73d6602ad3e02ecde7aac6190f1defe1d1d547e426f9afb7f72d2b23ba80b84c7e3bdd8d412a7580c483357992cba5607aaeabf78fd683eb9cd5The original Shorkie paper is the preferred scientific citation. If this
PyTorch port or method-rebuild weight is useful, you may additionally cite the
v0.1.0-rc1 software release.
Ziyan Zhuang
Tianjin University; Shenzhen Loop Area Institute
[email protected]
GitHub · Hugging Face
License: Apache-2.0.