---
license: apache-2.0
datasets:
- mlabonne/guanaco-llama2-1k
pipeline_tag: text-generation
---
# 🦙🧠 Miniguanaco-13b

📝 [Article](https://towardsdatascience.com/fine-tune-your-own-llama-2-model-in-a-colab-notebook-df9823a04a32) |
💻 [Colab](https://colab.research.google.com/drive/1PEQyJO1-f6j0S_XJ8DV50NkpzasXkrzd?usp=sharing) |
📄 [Script](https://gist.github.com/mlabonne/b5718e1b229ce6553564e3f56df72c5c)

<center><img src="https://i.imgur.com/1IZmjU4.png" width="300"></center>

This is a `Llama-2-13b-chat-hf` model fine-tuned using QLoRA (4-bit precision) on the [`mlabonne/guanaco-llama2-1k`](https://huggingface.co/datasets/mlabonne/guanaco-llama2-1k) dataset, which is a subset of the [`timdettmers/openassistant-guanaco`](https://huggingface.co/datasets/timdettmers/openassistant-guanaco).

## 🔧 Training

It was trained on an RTX 3090. It is mainly designed for educational purposes, not for inference. Parameters:

```
max_seq_length = 2048
use_nested_quant = True
bnb_4bit_compute_dtype=bfloat16
lora_r=8
lora_alpha=16
lora_dropout=0.05
per_device_train_batch_size=2
```

## 💻 Usage

``` python
# pip install transformers accelerate

from transformers import AutoTokenizer
import transformers
import torch

model = "mlabonne/llama-2-13b-miniguanaco"
prompt = "What is a large language model?"

tokenizer = AutoTokenizer.from_pretrained(model)
pipeline = transformers.pipeline(
    "text-generation",
    model=model,
    torch_dtype=torch.float16,
    device_map="auto",
)

sequences = pipeline(
    f'<s>[INST] {prompt} [/INST]',
    do_sample=True,
    top_k=10,
    num_return_sequences=1,
    eos_token_id=tokenizer.eos_token_id,
    max_length=200,
)
for seq in sequences:
    print(f"Result: {seq['generated_text']}")
```