neuva-lang 1.1.0

Neuva — A working ML programming language with lexer, parser, AST and interpreter

 ███╗   ██╗███████╗██╗   ██╗██╗   ██╗ █████╗
 ████╗  ██║██╔════╝██║   ██║██║   ██║██╔══██╗
 ██╔██╗ ██║█████╗  ██║   ██║██║   ██║███████║
 ██║╚██╗██║██╔══╝  ██║   ██║╚██╗ ██╔╝██╔══██║
 ██║ ╚████║███████╗╚██████╔╝ ╚████╔╝ ██║  ██║
 ╚═╝  ╚═══╝╚══════╝ ╚═════╝   ╚═══╝  ╚═╝  ╚═╝

A programming language built purely for Machine Learning.

Simple like Python. Clear like Rust. Made for ML.

---

What is Neuva?

Neuva (NOO-vah) is an open source programming language designed from the ground up for Machine Learning. No boilerplate. No imports. No configuration. Just describe your model and train it.

# Train a digit classifier — the entire program
let data = load("examples/data/iris.csv")
let train_data, test_data = data.split(0.8)

model DigitNet {
    layer dense(4 -> 16, relu)
    layer dense(16 -> 3,  softmax)
}

train DigitNet on train_data for 20 epochs, lr = 0.001, loss = crossentropy
print "Accuracy:", accuracy(DigitNet, test_data)

That is it. No import torch. No nn.Module. No training loop. Neuva handles it all.

---

What's New in v1.0.0

Neuva started as a single .lark grammar file and a stub interpreter. Over 34 days of public development — one commit per day — it grew into a complete ML language with a real PyTorch backend.

The journey in short:

• Days 5–9: grammar, parser, and first AST tests
• Days 10–11: working interpreter; Neuva ran its first program
• Days 12–17: real PyTorch backend; neural networks actually trained
• Days 20–21: real CSV data loading, train/test splits, 93% accuracy on Iris
• Days 22–25: VS Code extension, documentation, Rust-style error messages
• Days 27–30: else if chains, string concatenation, --version flag
• Days 31–33: interactive REPL, GPU detection, mini-batch training
• Day 34: f-strings, lists, and/or, dropout layers, model summary printing
• v1.0.0: 54 tests, 13 runnable examples, stable release

---

Why Neuva?

Most ML code looks like this:

import torch
import torch.nn as nn
import torch.optim as optim

class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.fc1 = nn.Linear(784, 128)
        self.fc2 = nn.Linear(128, 10)

    def forward(self, x):
        x = torch.relu(self.fc1(x))
        return torch.softmax(self.fc2(x), dim=1)

model = Net()
optimizer = optim.Adam(model.parameters(), lr=0.001)
criterion = nn.CrossEntropyLoss()

for epoch in range(20):
    for batch_x, batch_y in train_loader:
        optimizer.zero_grad()
        output = model(batch_x)
        loss = criterion(output, batch_y)
        loss.backward()
        optimizer.step()

Neuva code looks like this:

model DigitNet {
    layer dense(784 -> 128, relu)
    layer dense(128 -> 10,  softmax)
}

train DigitNet on train_data for 20 epochs, lr = 0.001

Same result. A fraction of the code.

---

Features

• ML-first keywords — model, layer, train, predict, save are built into the language
• No boilerplate — no imports, no class definitions, no training loops
• PyTorch backend — compiles to PyTorch under the hood, fast and compatible
• GPU support — automatically detects and uses CUDA if available
• Mini-batch training — shuffled batches each epoch, averaged loss reporting
• F-string interpolation — print "Accuracy: {acc}%" substitutes variables inline
• List support — let scores = [85, 90, 78], scores[0], len(scores)
• Boolean logic — and / or with short-circuit evaluation
• Dropout layers — layer dropout(0.3) maps directly to nn.Dropout
• Model summary — print MyModel shows layers and total parameter count
• Interactive REPL — neuva shell for live exploration with persistent state
• Rust-style errors — line numbers, column pointers, did-you-mean suggestions
• Simple types — tensor, matrix, int, float, bool, string
• Open source — AGPL v3 licensed, built in public, contributions welcome

---

Install

pip install neuva-lang

Run a .nva file:

neuva my_model.nva

Start the interactive shell:

neuva shell

Check the version:

neuva --version

---

Language Overview

Variables

let name   = "Neuva"        # string  (inferred)
let layers = 3              # int     (inferred)
let rate: float = 0.001     # float   (explicit)
let ready: bool = true      # bool    (explicit)

F-String Interpolation

let acc = 93
print "Accuracy: {acc}%"    # prints: Accuracy: 93%

let model_name = "DigitNet"
print "Training {model_name}..."

Lists

let scores = [85, 90, 78, 92]
print scores[0]             # 85
let n = len(scores)         # 4

Boolean Logic

if score > 50 and score < 100 {
    print "valid score"
}

if passed or bonus {
    print "you qualify"
}

Models

model MyNet {
    layer dense(784 -> 128, relu)
    layer dropout(0.3)
    layer dense(128 -> 10,  softmax)
}

print MyNet   # prints layer summary and parameter count

Supported layer types: dense, conv, pool, dropout, flatten

Supported activations: relu, sigmoid, softmax, tanh, linear

Loading Data

let data = load("examples/data/iris.csv")

let train_data, test_data = data.split(0.8)   # 80/20 split
data = data.normalize()                        # normalize values
data = data.shuffle()                          # shuffle rows

Training

# one-liner
train MyNet on train_data for 10 epochs, lr = 0.001

# multi-line (more readable)
train MyNet
    on    train_data
    for   50 epochs
    lr    = 0.0005
    loss  = crossentropy

Supported loss functions: crossentropy, mse, mae, binary_crossentropy

Predict and Evaluate

let acc = accuracy(MyNet, test_data)
print "Accuracy: {acc}"

Save and Load

save MyNet to "my_model.nva"
let loaded = load("my_model.nva")

Functions

fn welcome(name: string) {
    print "Hello from Neuva, {name}"
}

fn square(x: float) -> float {
    return x * x
}

Control Flow

if acc > 0.95 {
    print "Excellent model!"
} else if acc > 0.80 {
    print "Good model."
} else {
    print "Keep training."
}

for i in range(10) {
    print "Epoch {i}"
}

let count = 0
while count < 5 {
    let count = count + 1
}

Interactive REPL

$ neuva shell
Neuva 1.0.0 — interactive shell. Type 'exit' to quit.
>>> let x = 42
>>> print "The answer is {x}"
The answer is 42
>>> exit

---

Examples

File	Description
examples/hello.nva	Hello world
examples/digit_classifier.nva	Iris classification, 93% accuracy
examples/linear_regression.nva	House price regression
examples/spam_classifier.nva	Binary spam detection
examples/house_price.nva	Multi-layer regression
examples/cnn_classifier.nva	CNN layer definitions
examples/control_flow_demo.nva	if, for, while, functions
examples/elif_demo.nva	Chained else if
examples/fstring_demo.nva	F-string interpolation
examples/list_demo.nva	Lists and indexing
examples/logical_demo.nva	and / or operators
examples/dropout_demo.nva	Dropout layer and model summary
examples/string_concat_demo.nva	String concatenation

---

Project Status

Phase	Status	Description
1 — Design	✅ Done	Language syntax and keywords
2 — GitHub Setup	✅ Done	Repo, license, structure
3 — Lexer	✅ Done	Tokenizer with full terminal set
4 — Parser	✅ Done	Lark Earley parser, full AST
5 — Interpreter	✅ Done	Tree-walking interpreter, 54 tests
6 — PyTorch backend	✅ Done	Mini-batch training, GPU support
7 — PyPI release	✅ Done	pip install neuva-lang

---

Documentation

Document	Description
Getting Started	Install, hello world, running .nva files
Language Reference	Every keyword with syntax and examples
Examples	Line-by-line walkthrough of the iris classifier
Changelog	Full version history
Roadmap	What's planned post-1.0

---

Contributing

Neuva is open source and we welcome all contributors — beginners and experts alike.

git clone https://github.com/Ankit-Mahadani/neuva.git
cd neuva
pip install -e ".[dev]"
pytest tests/

See CONTRIBUTING.md for the full guide.

Look for issues labelled good first issue to get started: github.com/Ankit-Mahadani/neuva/issues

---

Built With

• Python 3.10+ — the interpreter is written in Python
• PyTorch ≥ 2.0 — ML execution backend
• Lark ≥ 1.1 — grammar and parser
• NumPy + Pandas — data loading and preprocessing

---

License

This project is licensed under the GNU Affero General Public License v3.0 (AGPL-3.0).

This means:

• You can use, study, modify, and distribute Neuva freely
• If you use Neuva in a product or service (including over a network), you must release your source code under the same license
• Any modifications must also be open source

See the full LICENSE file for details.

Copyright © 2026 Ankit Mahadani and Neuva Contributors

---

Built in public. Day by day. v1.0.0.

Star the repo to follow the journey.