2.9.26
Released 2 Nov 2026
Official Sigmora website
Neural network training agent for Windows
Sigmora is an agent for Windows that edits your Python training code, runs it and shows the real result of every step. Version 2.9.26 works on your own files and hardware.
Read the training log, find the error, and fix the shape mismatch in the model definition.
- read_file
train.pyloaded - grep
ValueError1 match - edit_file
train.pysaved - Shell
python train.pyexit 0 - read_file
loss.csv10 rows
waiting for your OK before the next stepRejectAccept
34 MB
Setup program download
Send feedback
Button inside the app
13
Interface languages
SHA-256
Verify installer hash
Auto
Update button shows
Neural network training on your own data
Neural network training means adjusting weights until the loss value falls on your training data. Your code does the work. Sigmora reads your scripts, edits them, runs your training command through Shell and shows every result as a card in the chat, and in Ask first mode you approve each change before it happens. Version 2.9.26 runs on Windows 10 and 11.
Fix a shape mismatch in neural network python code
A script stops. It fails with a shape error before the first epoch ends. In version 2.9.26, Sigmora uses grep to find the error line in your log, read_file to open train.py, and edit_file to change the layer size that does not match your labels. The diff appears as a card, and in Ask first mode you accept or reject it before the file changes. Shell then runs your training command and reports the real exit code, which only tells you that the script finished without an error.
Tool trace
grep ValueError → 1 match
read_file train.py → loaded
edit_file train.py → saved
Shell python train.py → exit 0Watch the loss in a backpropagation neural network run
You want the loss of every epoch in one file. Shell starts your training command in the environment you chose, and the real output lands in the chat. grep pulls the loss lines from the log, and write_file saves them as loss.csv in your working folder, so 2 runs can be compared later. In a neural network, each batch needs a forward pass and a backward pass, and the loss from the forward pass is what ends up in the CSV file. Ask first pauses there.
Tool trace
Shell python train.py → exit 0
grep loss → 10 matches
write_file loss.csv → savedChange the layers of a neural network architecture
You want a different layout of layers before the next run. edit_file proposes the new layer sizes in your model file and shows a diff, and in Ask first mode nothing is written until you accept it. Shell then runs a small batch through the model and prints the output shape. Compare it with your labels. If the 2 numbers differ, ask for another change. A passing run only shows that the script ran, so you still judge the design yourself.
Spot overfitting in a training experiment
Training loss keeps falling, but validation loss starts to rise. That is how to tell when your model is overfitting. Look at both columns. Sigmora uses grep to pull the 2 loss columns from your log, and edit_file shows the diff for a change you ask for, such as a smaller model or early stopping in your own script. In Ask first mode you approve it, then Shell reruns the script, and you compare validation loss before and after on data the model has not seen.
How Sigmora differs from a typical chat assistant
A chat assistant describes a command, while Sigmora 2.9.26 runs it and shows what really happened in the output card. You decide what runs.
| Typical chat assistant | Sigmora | |
|---|---|---|
| File editing | Writes code in the chat | Edits the file and shows a diff |
| Command execution | Describes a command | Runs it and returns the exit code |
| Approval mode | No approval step | Ask first pauses before changes |
| Model endpoint | Set by the chat service | Your endpoint, checked with Test |
How to get started with Sigmora
- Download the 34 MB setup program for Windows 10 or 11.
- Run the installer and open Sigmora.
- Connect your model endpoint or use the built-in model.
- Start a chat, name your training script and pick Auto-run, Ask first or Plan only.
Who uses Sigmora for training
Beginners writing their first network are the first audience. Which loss function works for a first project is a fair question, so the agent can run the same script with each choice and print the numbers side by side while you read them. Start with something small. Python has to be installed where the command runs.
Practitioners debugging training loops are the second audience. Shell runs your training command on your PC, in WSL or over SSH on your own GPU server, and returns the real output and exit code. You pick the environment. grep and read_file locate the failing line, and edit_file shows the diff first.
Researchers comparing experiments are the third audience. When you wonder what role the learning rate plays, the agent can rerun the script with 2 values and write both loss curves to a CSV file. Keep each run small. You read the results and decide what they mean on data the model has not seen.
Frequently asked questions
What is neural network training?
Training is the process of adjusting a network's weights so that the loss value, which measures how wrong its predictions are, becomes smaller on the training data. Each step runs a forward pass, a loss calculation, backpropagation and an optimiser update. Your Python code does this work, not Sigmora. Watch validation loss too.
Is neuralnetworkz.com the official Sigmora website?
Yes, neuralnetworkz.com is the official Sigmora website. The download button gives the setup program, about 34 MB, and records the time, IP address, browser user agent and referral source before it sends you to the installer. Compare the SHA-256 checksum with the published manifest. Check it first. Report problems with the Send feedback button in the app.
How does Sigmora handle privacy?
The chats, working folders and API keys stay on your PC, and secrets in tool output are redacted before they reach the model. Prompts, including the text of files the agent reads, go only to the model endpoint you pick. With a local model nothing leaves the machine. There is no telemetry.
What is backpropagation in a neural network?
Backpropagation is the step that computes how much each weight contributed to the loss, which gives the gradients. The optimiser then uses those gradients and the learning rate to change the weights. Sigmora does not do this itself; it runs your training script and shows the printed loss in a card in the chat. Read the numbers.
Does Sigmora work with local models?
Yes, you can connect a model that runs on your own PC through a local endpoint, and then nothing you send leaves the machine. With a cloud model your prompts go to that provider instead. The Test button checks an endpoint and shows its response time and context window. Press Test first.
What is a neural network tutorial?
A neural network tutorial is a step-by-step explanation of how to build and train a network, usually by writing the forward pass, the loss and the weight update yourself. Sigmora can write a small example file, run it and show the diff before the change is applied in Ask first mode. Read every line.
How do I verify the download?
The download is verified by running certutil -hashfile on the downloaded file with SHA256 and comparing the result with the SHA-256 value in the published manifest. The installer is not code-signed yet, so Windows SmartScreen may warn. Choose More info and Run anyway only after the checksum matches. Match first, then run.
Run your first training script with Sigmora
Sigmora 2.9.26 runs your Python scripts, edits your files and shows every result as a card in the chat; the 34 MB setup program is for Windows 10 or 11. Start small.
Download Sigmora for Windows ↗Explore more
Every page of this website is available in 20 languages.
Product features ↗ Model configuration ↗ Workflow examples ↗ Download installer ↗ FAQ ↗ Privacy policy ↗
