README.md

---
license: apache-2.0
datasets:
- tatsu-lab/alpaca
---

## 🍮 🦙 Flan-Alpaca: Instruction Tuning from Humans and Machines

📣 Curious to know the performance of 🍮 🦙 **Flan-Alpaca** on large-scale LLM evaluation benchmark, **InstructEval**? Read our paper [https://arxiv.org/pdf/2306.04757.pdf](https://arxiv.org/pdf/2306.04757.pdf). We evaluated more than 10 open-source instruction-tuned LLMs belonging to various LLM families including Pythia, LLaMA, T5, UL2, OPT, and Mosaic. Codes and datasets: [https://github.com/declare-lab/instruct-eval](https://github.com/declare-lab/instruct-eval)


Our [repository](https://github.com/declare-lab/flan-alpaca) contains code for extending the [Stanford Alpaca](https://github.com/tatsu-lab/stanford_alpaca)
synthetic instruction tuning to existing instruction-tuned models such as [Flan-T5](https://arxiv.org/abs/2210.11416).
The pretrained models and demos are available on HuggingFace 🤗 :

| Model                                                                     | Parameters | Training GPUs   |
|---------------------------------------------------------------------------|------------|-----------------|
| [Flan-Alpaca-Base](https://huggingface.co/declare-lab/flan-alpaca-base)   | 220M       | 1x A6000        |
| [Flan-Alpaca-Large](https://huggingface.co/declare-lab/flan-alpaca-large) | 770M       | 1x A6000        |
| [Flan-Alpaca-XL](https://huggingface.co/declare-lab/flan-alpaca-xl)       | 3B         | 1x A6000        |
| [Flan-Alpaca-XXL](https://huggingface.co/declare-lab/flan-alpaca-xxl)     | 11B        | 4x A6000 (FSDP) |

### Why?

[Alpaca](https://crfm.stanford.edu/2023/03/13/alpaca.html) represents an exciting new direction
to approximate the performance of large language models (LLMs) like ChatGPT cheaply and easily.
Concretely, they leverage an LLM such as GPT-3 to generate instructions as synthetic training data.
The synthetic data which covers more than 50k tasks can then be used to finetune a smaller model.
However, the original implementation is less accessible due to licensing constraints of the
underlying [LLaMA](https://ai.facebook.com/blog/large-language-model-llama-meta-ai/) model.
Furthermore, users have noted [potential noise](https://github.com/tloen/alpaca-lora/issues/65) in the synthetic
dataset. Hence, it may be better to explore a fully accessible model that is already trained on high-quality (but
less diverse) instructions such as [Flan-T5](https://arxiv.org/abs/2210.11416).

### Usage
This uses Huggingface PEFT library for Parameter Efficient Fine Tuning

```
import torch
from peft import PeftModel
from transformers import GenerationConfig


from transformers import AutoTokenizer, AutoModelForSeq2SeqLM


BASE_MODEL = "google/flan-t5-xl"
LORA_WEIGHTS = "declare-lab/flan-alpaca-xl-lora"
TEMPERATURE = 1.0
TOP_P = 0.75
TOP_K = 40
NUM_BEAMS = 4
MAX_NEW_TOKENS = 128

if torch.cuda.is_available():
    device = "cuda"
else:
    device = "cpu"


if device == "cuda":
    model = AutoModelForSeq2SeqLM.from_pretrained(
        BASE_MODEL,
        device_map="auto",
    )
    model = PeftModel.from_pretrained(model, LORA_WEIGHTS, force_download=True)
else:
    model = AutoModelForSeq2SeqLM.from_pretrained(
        BASE_MODEL, device_map={"": device}, low_cpu_mem_usage=True
    )
    model = PeftModel.from_pretrained(
        model,
        LORA_WEIGHTS,
        device_map={"": device},
    )


prompt = "Write a short email to show that 42 is the optimal seed for training neural networks"

tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)
input_ids = tokenizer(prompt, return_tensors="pt").input_ids
input_ids = input_ids.to(device)

generation_config = GenerationConfig(
    temperature=TEMPERATURE,
    top_p=TOP_P,
    top_k=TOP_K,
    num_beams=NUM_BEAMS,
)
generation_output = model.generate(
    input_ids=input_ids,
    generation_config=generation_config,
    return_dict_in_generate=True,
    output_scores=True,
    max_new_tokens=MAX_NEW_TOKENS,
)
print(tokenizer.batch_decode(generation_output.sequences)[0])

```