AtomLlama-33K-5x5-DigitMesh-Sparse

A 50% sparse version of atomllama-33K-5x5-DigitMesh for efficient 5×5 digit mesh recognition.

Model Description

This is a ~50% unstructured sparse variant of the AtomLlama-33K-5x5-DigitMesh model, pruned using the SparseGPT algorithm. Half of the model weights have been set to zero while maintaining digit recognition accuracy through second-order optimization.

Key Features

• Base Model: junzzhu/atomllama-33K-5x5-DigitMesh
• Sparsity: ~50% (unstructured)
• Pruning Method: SparseGPT with Hessian-based importance scoring
• Parameters: ~33K total, ~16.5K non-zero
• Architecture: LlamaForCausalLM
• Task: 5×5 binary digit mesh recognition

Usage

Serving with vLLM

python -m vllm.entrypoints.openai.api_server \
  --model junzzhu/atomllama-33K-5x5-DigitMesh-sparse \
  --max-model-len 32

Example Inference

curl http://localhost:8000/v1/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "junzzhu/atomllama-33K-5x5-DigitMesh-sparse",
    "prompt": "1 1 1 1 1 1 0 0 0 1 1 0 0 0 1 1 0 0 0 1 1 1 1 1 1 <SEP>",
    "max_tokens": 1,
    "temperature": 0
  }'

Expected output: D0

Sparsity Details

• Type: Unstructured (weights pruned individually based on importance)
• Target Sparsity: 50%
• Calibration: Pruned using digit pattern activations
• Benefits: Reduced memory footprint and potential inference speedup with sparse tensor libraries

License

Apache-2.0

Citation

@misc{atomllama-33k-digitMesh-sparse,
  title={AtomLlama-33K-5x5-DigitMesh-Sparse: A 50% Sparse Model for Digit Recognition},
  author={Jun Zhu},
  year={2026},
  howpublished={\url{https://huggingface.co/junzzhu/atomllama-33K-5x5-DigitMesh-sparse}}
}