---
base_model: neuralmagic/Llama-2-7b-pruned70-retrained
inference: true
model_type: llama
pipeline_tag: text-generation
datasets:
  - cerebras/SlimPajama-627B
  - HuggingFaceH4/ultrachat_200k
tags:
- sparse
- chat
---

# Llama-2-7b-pruned70-retrained-ultrachat

This repo contains a [70% sparse Llama 2 7B](https://huggingface.co/neuralmagic/Llama-2-7b-pruned70-retrained) finetuned for chat tasks using the [UltraChat 200k](https://huggingface.co/datasets/HuggingFaceH4/ultrachat_200k) dataset.

**Authors**: Neural Magic, Cerebras

## Usage

Below we share some code snippets on how to get quickly started with running the model.

### Sparse Transfer

By leveraging a pre-sparsified model's structure, you can efficiently fine-tune on new data, leading to reduced hyperparameter tuning, training times, and computational costs. Learn about this process [here](https://neuralmagic.github.io/docs-v2/get-started/transfer).

### Running the model

This model may be run with the transformers library. For accelerated inference with sparsity, deploy with [nm-vllm](https://github.com/neuralmagic/nm-vllm) or [deepsparse](https://github.com/neuralmagic/deepsparse).

```python
# pip install transformers accelerate
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("neuralmagic/Llama-2-7b-pruned70-retrained-ultrachat")
model = AutoModelForCausalLM.from_pretrained("neuralmagic/Llama-2-7b-pruned70-retrained-ultrachat", device_map="auto")

input_text = "Write me a poem about Machine Learning."
input_ids = tokenizer.apply_chat_template(input_text, add_generation_prompt=True, return_tensors="pt").to("cuda")

outputs = model.generate(**input_ids)
print(tokenizer.decode(outputs[0]))
```

## Evaluation Benchmark Results

Model evaluation metrics and results.

| Benchmark                                      | Metric        | Llama-2-7b-ultrachat  | Llama-2-7b-pruned70-retrained-ultrachat |
|------------------------------------------------|---------------|-------------|-------------------------------|
| [MMLU](https://arxiv.org/abs/2009.03300)       | 5-shot        | 46.1%       | 32.5%                         |
| [HellaSwag](https://arxiv.org/abs/1905.07830)  | 0-shot        | 75.9%       | 68.9%                         |
| [WinoGrande](https://arxiv.org/abs/1907.10641) | 5-shot        | 72.6%       | 65.1%                         |
| [ARC-c](https://arxiv.org/abs/1911.01547)      | 25-shot       | 52.8%       | 45.3%                         |
| [TruthfulQA](https://arxiv.org/abs/2109.07958) | 5-shot        | 44.8%       | 39.6%                         |
| [GSM8K](https://arxiv.org/abs/2110.14168)      | 5-shot        | 12.4%       | 4.8%                          |
| [AlpacaEval](https://arxiv.org/abs/2107.03374) ([Llama-2-7b-chat-hf](https://huggingface.co/meta-llama/Llama-2-70b-chat-hf) evaluator) | Win rate  | 57.6% | 57.4% |
| [AlpacaEval](https://arxiv.org/abs/2107.03374) (GPT-4 Turbo evaluator) | Win rate  | 60.6% | 54.0% |

## Model Training Details

This model was obtained by sparse-tranfer of the sparse foundational model [Llama-2-7b-pruned70-retrained](https://huggingface.co/neuralmagic/Llama-2-7b-pruned70-retrained) on the [ultrachat_200k](https://huggingface.co/datasets/HuggingFaceH4/ultrachat_200k) dataset.
Training was performed for 2 epochs and used the [SquareHead](https://arxiv.org/abs/2310.06927) knowledge distillation with [Llama-2-7b-ultrachat](https://huggingface.co/neuralmagic/Llama-2-7b-ultrachat) as teacher.

## Help

For further support, and discussions on these models and AI in general, join [Neural Magic's Slack Community](https://join.slack.com/t/discuss-neuralmagic/shared_invite/zt-q1a1cnvo-YBoICSIw3L1dmQpjBeDurQ)