Instructions to use ameythakur/SAIR-Modular-Arithmetic-Challenge with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use ameythakur/SAIR-Modular-Arithmetic-Challenge with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="ameythakur/SAIR-Modular-Arithmetic-Challenge")

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("ameythakur/SAIR-Modular-Arithmetic-Challenge", dtype="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use ameythakur/SAIR-Modular-Arithmetic-Challenge with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "ameythakur/SAIR-Modular-Arithmetic-Challenge"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "ameythakur/SAIR-Modular-Arithmetic-Challenge",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/ameythakur/SAIR-Modular-Arithmetic-Challenge

SGLang

How to use ameythakur/SAIR-Modular-Arithmetic-Challenge with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "ameythakur/SAIR-Modular-Arithmetic-Challenge" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "ameythakur/SAIR-Modular-Arithmetic-Challenge",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "ameythakur/SAIR-Modular-Arithmetic-Challenge" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "ameythakur/SAIR-Modular-Arithmetic-Challenge",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use ameythakur/SAIR-Modular-Arithmetic-Challenge with Docker Model Runner:
```
docker model run hf.co/ameythakur/SAIR-Modular-Arithmetic-Challenge
```

SAIR Modular Arithmetic Challenge: Abacus Generalization

Overview

This repository hosts the weights for the SAIR Modular Arithmetic Challenge. The model is an autoregressive transformer built to solve exact modular multiplication (A × B) mod P organically, without hardcoded arithmetic logic, symbolic parsers, or external computation modules.

Using state-machine decoding, this research targets the learnability wall of transformer mathematics, achieving theoretical infinite length generalization.

Architectural Methods

To bypass the spatial limitations of standard transformers, this model uses three methods:

1. Abacus Significance Embeddings

Standard transformers track coordinate positions. This model strips coordinate embeddings, replacing them with Mathematical Significance Injections. Digits route based on their place-value, ensuring 1024-bit primes process through the exact same logic gates as 16-bit primes.

2. Algorithmic Scratchpads (Bit-Serial Decoding)

The network operates as a recurrent state machine. By forcing the model to generate intermediate computational traces autoregressively, the network allocates computation proportionally to integer complexity, mimicking a Turing machine tape.

3. Grokking Phase Transitions

The weights deployed here were captured after the grokking phase transition. The model trained through thousands of delayed gradient steps beyond the initial validation plateau with extreme weight decay, forcing the network to collapse memorization circuits into sparse mathematical algorithms.

Inference Format

This model requires a specific execution format. If prompted with an equation, it sequentially emits the step-by-step logic trace before terminating natively with the final matrix node.

Input Format: Character-level ASCII equations.

123*456

Output Format: State machine algorithmic trace terminating in ANS=.

(scratchpad logic...) ANS=56088<EOS>

Technical Specifications

Architecture Layer: Bit-Serial Autoregressive Transformer
Embedding Mechanism: Significance Input Injection (Abacus)
Tokenization: Discrete Character-level (Base10Tokenizer)
Framework Ecosystem: PyTorch 2.0+

Citations and Laboratory

This model is an isolated artifact. The complete research laboratory - containing the synthetic data generators, the training loops, the sandbox validators, and the PyTorch implementations - is open-source.

Official Research Repository:
SAIR-MODULAR-ARITHMETIC-CHALLENGE

Prepared by Amey Thakur for the SAIR Foundation AI Benchmark.

Downloads last month: -; Downloads are not tracked for this model. How to track

Safetensors

Model size

671k params

Tensor type

F32