---
license: mit
datasets:
- vesteinn/babylm
---
# rootxhacker/arthemis-lm

Building capable language models shouldn't require massive corporate budgets. While the industry pushes toward increasingly large models, this project explores what's possible with neuromorphic architectures and limited resources.

I developed this 155.8M parameter Llama-SNN-LTC model with specific constraints:

- Budget limit: Under $50 using Google Colab Pro Plus
- From-scratch pretraining with fully open-source dataset  
- No fine-tuning or synthetic data generation from existing LLMs
- Focus on architectural innovation over scale

## Model Details

This project incorporates **Spiking Neural Networks (SNNs)** and **Liquid Time Constants (LTCs)** into the Llama architecture, creating a neuromorphic language model. I spent under $50 on Google Colab Pro Plus and used the first 1M samples from the BabyLM challenge dataset, which contains approximately 100M tokens.
This model is working on par with google/bert-large-uncased model

**Model Type**: Causal Language Model with Neuromorphic Enhancements  
**Supported Languages**: English  
**Number of Parameters**: 155.8M  
**Context Length**: 1024 tokens  
**Base Architecture**: Llama with SNN/LTC modifications  
**Training Data**: BabyLM (vesteinn/babylm) - 1M samples (~100M tokens)

### Architecture Features
- **Spiking Neural Networks** in attention mechanisms for temporal processing
- **Liquid Time Constants** in feed-forward layers for adaptive dynamics
- **12-layer transformer backbone** with neuromorphic enhancements
- **RoPE positional encoding** for sequence understanding
- **Custom surrogate gradient training** for differentiable spike computation

Here are my major model configurations:

```
hidden_size = 768
intermediate_size = 2048
num_hidden_layers = 12
num_attention_heads = 12
num_key_value_heads = 12
max_position_embeddings = 1024
vocab_size = 50257
spiking_threshold = 1.0
ltc_hidden_size = 256
ltc_layers = 2
```

## Usage

### Install dependencies
```bash
pip install transformers torch numpy
```

## Inference
This gist has full code for inference

``` bash
https://gist.github.com/harishsg993010/e632de8b15a3ab1ff03e3912f55109ea
```

## Evaluation

I performed evaluation using https://gist.github.com/harishsg993010/e3c31c2d2c8207384ee263627f990300

### Results Comparison

| Model | Params | Budget | HellaSwag | OBQA | WinoGrande | ARC_e | ARC_c | BoolQ | Avg |
|-------|--------|--------|-----------|------|------------|-------|-------|-------|-----|
| **rootxhacker/arthemis-lm** | **155.8M** | **<$50** | **24.65** | **20.60** | **48.10** | **28.20** | **22.20** | **39.80** | **30.59** |
| google/bert-large-uncased | 336M | N/A | 24.53 | 26.20 | 49.80 | 25.08 | 25.68 | 40.86 | 32.03 |

## Observations

- **Budget Efficiency**: Our model achieves competitive performance with only ~$50 budget, demonstrating that meaningful language models can be built with limited resources.
- **Neuromorphic Advantages**: The SNN-LTC architecture shows particularly strong performance in WinoGrande (48.10%), suggesting enhanced reasoning capabilities from temporal dynamics.
- **Parameter Efficiency**: With 155.8M parameters, our model performs comparably to BERT-large-uncased (336M parameters) while being significantly smaller.
- **Room for Improvement**: More training data and compute would likely improve performance, but the current results validate the neuromorphic approach.



```
Architecture: Llama + Spiking Neural Networks + Liquid Time Constants
Hidden Size: 768
Intermediate Size: 2048
Attention Heads: 12
Layers: 12
Max Position Embeddings: 1024
Vocabulary Size: 50,257
Spiking Threshold: 1.0
LTC Hidden Size: 256
Training Precision: FP32
```

## Training Details

The model was pretrained from scratch using:
- **Dataset**: BabyLM (vesteinn/babylm) - First 1M samples (~100M tokens)
- **Hardware**: Google Colab Pro Plus (A100 GPU)
- **Training Steps**: 20,000 steps
- **Batch Size**: 8 with gradient accumulation
- **Learning Rate**: 3e-4 with linear warmup
- **Precision**: FP32 for stability with neuromorphic components

### Key Innovations
- **Custom SNN Implementation**: Leaky Integrate-and-Fire neurons with surrogate gradients
- **Liquid Time Constants**: Adaptive time dynamics in feed-forward layers
- **Budget-Conscious Training**: Optimized for maximum performance per dollar spent
- **Neuromorphic Language Modeling**: First known integration of SNNs and LTCs in causal LM

## Future Work

- Scale to larger datasets with increased compute budget
- Explore different spiking neuron models (e.g., Adaptive LIF, Izhikevich)
- Implement more sophisticated LTC architectures
- Fine-tune for specific downstream tasks
- Compare energy efficiency with standard transformers

## Model Sources

- **Repository**: [Coming Soon]
- **Paper**: [In Progress]
- **Hugging Face**: [rootxhacker/arthemis-lm](https://huggingface.co/rootxhacker/arthemis-lm)

## Uses

This model can be used for:
- Text generation and completion
- Few-shot learning tasks
- Research into neuromorphic language models
- Educational purposes for understanding SNN/LTC architectures
- Base model for fine-tuning on specific tasks

## Limitations

- **Training Data**: Limited to 100M tokens (much smaller than typical LLMs)
- **Context Length**: Maximum 1024 tokens
- **Domain**: Primarily trained on English text
- **Compute**: Training limited by budget constraints
- **Performance**: Lower than larger, more extensively trained models

## Acknowledgments

Special thanks to **keeeeenw** for the inspiration and open-source MicroLlama project, which demonstrated that impressive language models can be built on a budget. This work builds upon those principles while exploring neuromorphic computing approaches to language modeling.

## Citation

```bibtex
@misc{arthemis-lm-2024,
  title={Arthemis-LM: A Neuromorphic Language Model with Spiking Neural Networks and Liquid Time Constants},
  author={rootxhacker},
  year={2024},
  howpublished={\url{https://huggingface.co/rootxhacker/arthemis-lm}}
}
```

## License

Apache License 2.0