README.md

---
datasets:
- HuggingFaceH4/ultrachat_200k
language:
- en
pipeline_tag: text-generation
---

# SparseLlama-2-7b-ultrachat_200k-pruned_50.2of4

## Model Overview
- **Model Architecture:** Llama-2
  - **Input:** Text
  - **Output:** Text
- **Model Optimizations:**
  - **Pruned:** 50% 2:4
- **Release Date:** 6/28/2024
- **Version:** 1.0
- **Model Developers:** Neural Magic

Compressed version of [Llama-2-7b](https://huggingface.co/meta-llama/Llama-2-7b-hf) specialized for text-generation.
This model was obtained by fine-tuning the Sparse Foundational model [Sparse-Llama-2-7b-pruned_50.2of4](https://huggingface.co/nm-testing/SparseLlama-2-7b-pruned_50.2of4) on the [ultrachat_200k](https://huggingface.co/datasets/HuggingFaceH4/ultrachat_200k) dataset, using [SquareHead distillation] (https://arxiv.org/abs/2310.06927) and [Llama-2-7b-ultrachat200k](https://huggingface.co/neuralmagic/Llama-2-7b-ultrachat200k) as teacher.
It achieves a win rate of 64.9% on the [AlpacaEval](https://github.com/tatsu-lab/alpaca_eval) benchmark (version 1.0) when using [Llama-2-70b-chat](https://huggingface.co/meta-llama/Llama-2-70b-chat-hf) as evaluator, whereas the dense [Llama-2-7b-ultrachat200k](https://huggingface.co/neuralmagic/Llama-2-7b-ultrachat200k) model achieves 57.6% win rate.

This model was produced as part if Neural Magic's Sparse Foundational Models initiative, and demostrates the capability of Sparse Foundational Models to transfer to the text-generation domain.

**Note:** This model uses the chat template from [zephyr-7b-beta](https://huggingface.co/HuggingFaceH4/zephyr-7b-beta).

## Model Optimizations

This model is derived from the Sparse Foundational model [Sparse-Llama-2-7b-pruned_50.2of4](https://huggingface.co/nm-testing/SparseLlama-2-7b-pruned_50.2of4), which was obtained by applying the [SparseGPT](https://arxiv.org/abs/2301.00774) algorithm to prune [Llama-2-7b](https://huggingface.co/meta-llama/Llama-2-7b-hf) to 50% sparsity with a 2:4 mask.
This optimization reduces the number of parameters by 50%, reducing the disk size and FLOPs by the same level.

## Evaluation

This model was evaluated in the [AlpacaEval](https://github.com/tatsu-lab/alpaca_eval) benchmark using [Llama-2-70b-chat](https://huggingface.co/meta-llama/Llama-2-70b-chat-hf) as evaluator.

## Accuracy

| Model | Win rate | Recovery |
| :----- | :--------: | :--------: |
| [Llama-2-7b](https://huggingface.co/meta-llama/Llama-2-7b-hf) | 3.7% | -- |
| [Llama-2-7b-ultrachat200k](https://huggingface.co/neuralmagic/Llama-2-7b-ultrachat200k) | 57.6% | -- |
| SparseLlama-2-7b-ultrachat_200k-pruned_50.2of4 | 64.9% | 113% |