Llama-2-7b-pruned70-retrained-instruct
This repo contains a 70% sparse Llama 2 7B finetuned for instruction-following tasks using a blend of the Platypus + Open Orca + Dolphin datasets.
Official model weights from Enabling High-Sparsity Foundational Llama Models with Efficient Pretraining and Deployment.
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("Llama-2-7b-pruned70-retrained-instruct")
model = AutoModelForCausalLM.from_pretrained("Llama-2-7b-pruned70-retrained-instruct", device_map="auto")
input_text = "Write a recipe for banana bread:\n"
input_ids = tokenizer(input_text, 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-instruct | Llama-2-7b-pruned70-retrained-instruct |
|---|---|---|---|
| MMLU | 5-shot | 48.60% | 42.33% |
| HellaSwag | 10-shot | 79.45% | 77.21% |
| WinoGrande | 5-shot | 75.69% | 71.90% |
| ARC-c | 25-shot | 53.92% | 47.35% |
| TruthfulQA | 0-shot | 43.63% | 42.25% |
| GSM8K | 5-shot | 15.92% | 14.25% |
Model Training Details
This model was obtained by sparse-tranfer of the sparse foundational model Llama-2-7b-pruned50-retrained on a blend of Open Platypus, 10% Open Orca and 10% Dolphin datasets. Training was perfomerd for 6 epochs.
Help
For further support, and discussions on these models and AI in general, join Neural Magic's Slack Community