---
license: apache-2.0
---
# BrainTransformers: SNN-LLM

Based on BrainTransformers, BrainGPTForCausalLM is a Large Language Model (LLM) implemented using Spiking Neural Networks (SNN). Our technical report will be uploaded to arXiv as soon as possible. We plan to further optimize the model at the operator level and adapt it for hardware compatibility, enabling BrainGPTForCausalLM to be deployed on more energy-efficient SNN hardware devices.

## Model Availability

- The current pre-trained model parameters have been published on ModelScope: [DataLinguistic/BrainTransformers-3B-Chat](https://www.modelscope.cn/models/DataLinguistic/BrainTransformers-3B-Chat)
- The current pre-trained model parameters have been published on Hugging Face.[LumenscopeAI/BrainTransformers-3B-Chat](https://huggingface.co/LumenscopeAI/BrainTransformers-3B-Chat)

## Repository

The github link is: [LumenScopeAI/BrainTransformers-SNN-LLM](https://github.com/LumenScopeAI/BrainTransformers-SNN-LLM)

## Model Performance

Below are the performance metrics of our 3B model on various benchmarks:

| Task Category | Dataset | Performance |
|---------------|---------|-------------|
| General Tasks | MMLU | 65.6 |
|               | MMLU-pro | 34.6 |
|               | MMLU-redux | 63.7 |
|               | BBH | 56.3 |
|               | ARC-C | 56.5 |
|               | Trurhfulqa | 48.9 |
|               | Winogrande | 71.1 |
|               | Hellaswag | 74.6 |
| Math and Science Tasks | GPQA | 26.3 |
|                        | Theoremqa | 27.4 |
|                        | MATH | 42.6 |
|                        | MMLU-stem | 62.5 |
|                        | GSM8K | 79.1 |
| Coding Tasks | HumanEval | 42.1 |
|              | HumanEval+ | 36.0 |
|              | MBPP | 57.1 |
|              | MBPP+ | 49.4 |
|              | MultiPL-E | 41.2 |
| Multilingual Tasks | Multi-Exam | 54.6 |
|                    | Multi-Understanding | 76.6 |
|                    | Multi-Mathematics | 48.9 |
|                    | Multi-Translation | 29.3 |

## Usage

### Generate Text
```python
import torch
from transformers import AutoTokenizer, BrainGPTForCausalLM

model_path = "/path/to/your/model"
model = BrainGPTForCausalLM.from_pretrained(model_path)
tokenizer = AutoTokenizer.from_pretrained(model_path)

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)

def generate_text(messages, max_new_tokens=50):
    text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    model_inputs = tokenizer([text], return_tensors="pt").to(device)
    
    with torch.no_grad():
        generated_ids = model.generate(**model_inputs, max_new_tokens=max_new_tokens)
    
    generated_ids = [output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)]
    return tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]

# Example usage
messages = [
    {"role": "system", "content": "You are a knowledgeable assistant."},
    {"role": "user", "content": "Explain the Pythagorean theorem."}
]
response = generate_text(messages)
print(response)
```

---
model-index:
  - name: BrainTransformers-3B-Chat
    results:
      - task:
          type: text-generation
        dataset:
          name: mmlu
          type: mmlu
        metrics:
          - name: MMLU
            type: MMLU
            value: 65.6
      - task:
          type: text-generation
        dataset:
          name: bbh
          type: bbh
        metrics:
          - name: BBH
            type: BBH
            value: 56.3
      - task:
          type: text-generation
        dataset:
          name: arc-challenge
          type: arc-challenge
        metrics:
          - name: ARC-C
            type: ARC-C
            value: 56.5
      - task:
          type: text-generation
        dataset:
          name: hellaswag
          type: hellaswag
        metrics:
          - name: HellaSwag
            type: HellaSwag
            value: 74.6
      - task:
          type: text-generation
        dataset:
          name: gsm8k
          type: gsm8k
        metrics:
          - name: GSM8K
            type: GSM8K
            value: 79.1
      - task:
          type: code-generation
        dataset:
          name: humaneval
          type: humaneval
        metrics:
          - name: HumanEval
            type: HumanEval
            value: 42.1
    source:
      name: LumenScopeAI
      url: https://github.com/LumenScopeAI/BrainTransformers-SNN-LLM
---