---
license: apache-2.0
language:
- en
library_name: transformers
tags:
- salienceformer
- memory
- hippocampus
- neuroscience
- gemma
- language-model
base_model: google/gemma-2b
datasets:
- wikitext
metrics:
- perplexity
pipeline_tag: text-generation
---

# SalienceFormer-Gemma2B

<p align="center">
  <img src="https://img.shields.io/badge/PyTorch-2.0+-red?logo=pytorch" alt="PyTorch">
  <img src="https://img.shields.io/badge/Transformers-4.36+-yellow?logo=huggingface" alt="Transformers">
  <img src="https://img.shields.io/badge/Base-Gemma--2B-blue" alt="Gemma">
</p>

**SalienceFormer** is a biologically-inspired memory architecture that brings hippocampal memory consolidation to large language models. This model integrates hippocampal mechanisms directly into the Gemma-2B transformer.

## Model Description

SalienceFormer adds three key components inspired by how the human hippocampus processes memories:

| Component | Inspiration | Function |
|-----------|-------------|----------|
| **Salience Gate** | Sharp Wave Ripples (SPW-Rs) | Dual-pathway importance scoring |
| **Memory Buffer** | Sleep Replay | Priority-based consolidation |
| **Drift Calibrator** | Synaptic Homeostasis | Embedding stability |

### Architecture

```
Input Tokens -> Gemma-2B (frozen + LoRA) -> Hidden States
    -> Salience Gate (importance scoring)
    -> Drift Calibrator (stability)
    -> Memory Buffer (consolidation)
    -> Output Fusion (cross-attention)
    -> Output Logits
```

## Results

### Perplexity (WikiText-2)

| Model | Parameters | Perplexity |
|-------|------------|------------|
| GPT-2 | 124M | 29.41 |
| Gemma-2B | 2B | ~18 |
| **SalienceFormer** | 2B + 15M | **11.83** |

### Ablation Study

| Configuration | PPL | Impact |
|--------------|-----|--------|
| Full SalienceFormer | 11.83 | baseline |
| No Salience Gate | 39.75 | +27.92 |
| No Memory Buffer | 89.84 | +78.01 |

### Brain-Like Behavior

| Metric | Value | Interpretation |
|--------|-------|----------------|
| Content/Function Ratio | 2.11x | Selective memory (content words tagged more) |
| Long-Range Benefit | +6.95 PPL | Better context retention |
| Buffer Priority | 4.9/5.0 | High-importance retention |

## Usage

```python
from salienceformer import SalienceFormer, SalienceFormerConfig
from huggingface_hub import hf_hub_download
import torch

# Download checkpoint
ckpt_path = hf_hub_download(
    repo_id="Gustav-Proxi/SalienceFormer-Gemma2B",
    filename="pytorch_model.pt"
)

# Initialize model
config = SalienceFormerConfig(
    base_model_name="google/gemma-2b",
    freeze_base=True,
    use_lora=True,
)
model = SalienceFormer(config)

# Load weights
ckpt = torch.load(ckpt_path, map_location="cpu")
model.load_state_dict(ckpt["model_state_dict"], strict=False)

# Generate
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("google/gemma-2b")
inputs = tokenizer("The capital of France is", return_tensors="pt")
outputs = model.generate(inputs["input_ids"], max_new_tokens=20)
print(tokenizer.decode(outputs[0]))
```

## Training

- **Base Model**: Gemma-2B (frozen with LoRA)
- **Dataset**: WikiText-2
- **Hardware**: NVIDIA RTX 4090 (24GB)
- **Training Time**: ~24 hours
- **Best Checkpoint**: step-110000

## Citation

```bibtex
@misc{salienceformer2025,
  title={SalienceFormer: Salience-Gated Memory Consolidation for Transformers},
  author={Vaishak Girish Kumar and Sanika},
  year={2025},
  howpublished={\url{https://github.com/Gustav-Proxi/SalienceFormer}},
}
```

## Links

- **GitHub**: [Gustav-Proxi/SalienceFormer](https://github.com/Gustav-Proxi/SalienceFormer)
- **Paper**: Coming soon

## License

Apache 2.0