Llama-2-7b-pruned50-retrained-ultrachat

This repo contains a 50% sparse Llama 2 7B finetuned for chat tasks using the 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.

Running the model

This model may be run with the transformers library. For accelerated inference with sparsity, deploy with nm-vllm or deepsparse.

# pip install transformers accelerate
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("neuralmagic/Llama-2-7b-pruned50-retrained-ultrachat")
model = AutoModelForCausalLM.from_pretrained("neuralmagic/Llama-2-7b-pruned50-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-pruned50-retrained-ultrachat
MMLU	5-shot	46.1%	41.4%
HellaSwag	0-shot	75.9%	73.5%
WinoGrande	5-shot	72.6%	67.8%
ARC-c	25-shot	52.8%	49.0%
TruthfulQA	5-shot	44.8%	39.5%
GSM8K	5-shot	12.4%	8.0%
AlpacaEval (Llama-2-7b-chat-hf evaluator)	Win rate	57.6%	60.1%
AlpacaEval (GPT-4 Turbo evaluator)	Win rate	60.6%	59.0%

Model Training Details

This model was obtained by sparse-tranfer of the sparse foundational model Llama-2-7b-pruned50-retrained on the ultrachat_200k dataset. Training was perfomerd for 2 epochs and used the SquareHead knowledge distillation with 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