---
license: mit
datasets:
- Reza8848/MUFFIN_68k
---


<img src="https://cdn-uploads.huggingface.co/production/uploads/6434a6e8ea46c009904c617e/J_4FHXmtM6TuRnN3aL06y.png" width="38" height="38">

This is the Llama2 LoRA weight that was fine-tuned on **MUFFIN** (**Mu**lti-**F**aceted **In**structions).

We fine-tune the [Llama2-13B](https://huggingface.co/meta-llama/Llama-2-13b-hf) on [MUFFIN dataset](https://arxiv.org/abs/2312.02436) with LoRA (low-rank adaption).

We released the LoRA weights of both Llama2 7B and 13B models:
|Model|LoRA Target Modules|
|-|-|
|[MUFFIN-Llama2-7B](https://huggingface.co/Reza8848/MUFFIN-Llama2-lora-7B)|`Q, K, V, O`|
|[MUFFIN-Llama2-13B](https://huggingface.co/Reza8848/MUFFIN-Llama2-lora-13B)|`Q, K, V, O`|

You can also find the T5-based models [here](https://huggingface.co/Reza8848/MUFFIN-T5-3B).


## Model Usage

### 1. Inference code

We use [Alpaca-lora](https://github.com/tloen/alpaca-lora) as our fine-tuning code.

So, when adopting the released model weights for inference, it should be better to use the [generation code](https://github.com/tloen/alpaca-lora/blob/main/generate.py) of Alpaca-lora to reproduce our performance.


Please follow the document of Alpaca-lora to set up the **correct Python environments first**.


> Our released lora weights are in **`.safetensors`** format rather than the common **`.bin`** torch model files.
> Wrong transformers and torch versions may result in [PEFT compatibility errors](https://github.com/huggingface/transformers/issues/27397) when using the released lora weighs.


### 2. Prompt template

Please use the following prompt template (save the following dict as a JSON file under ['template' folder](https://github.com/tloen/alpaca-lora/tree/main/templates)):

```json
{
    "description": "Template used by muffin.",
    "prompt_input": "### Input:\n{input}\n\n### Instruction:\n{instruction}\n\n### Response:\n",
    "prompt_no_input": "### Input:\nNone\n\n### Instruction:\n{instruction}\n\n### Response:\n",
    "response_split": "### Response:"    
}
```

### 3. Generation hyper-parameters

We use the default generation hyper-parameters as identified in [this line](https://github.com/tloen/alpaca-lora/blob/main/generate.py#L90).

Besides, be aware of the following hyper-parameters:
- `max_input_len == 1024`. This is the max_input_len of training. But it's fine to use any length in the inference since our evaluation batch size is 1.
- `num_beams == 1`. In our experiments, we set beam size to 1. But we recommend you try with a larger beam size to get better responses from models.
- When doing batched inference, please make sure `tokenizer.padding_side = "left" `, as we left padded all the batched instances when doing tuning (though it shall not have a big impact on the inference results).

## Zero-Shot Evaluation Performances

We use the [metric calculation scripts](https://github.com/yizhongw/Tk-Instruct/blob/main/src/compute_metrics.py) of [Tk-Instruct](https://github.com/yizhongw/Tk-Instruct/tree/main) (i.e., `ROUGE-L` and `Exact-Match`). 

<div style="text-align:center"><img src="https://cdn-uploads.huggingface.co/production/uploads/6434a6e8ea46c009904c617e/IjeMYWLMRO_qGOOiXxemP.png" alt="performances.png" width="600"/></div>



## 🥳 Citation

Please kindly cite our paper if you use any resources in this repository:

```bibtex
@inproceedings{Lou2023MUFFIN,
   title={{MUFFIN}: Curating Multi-Faceted Instructions for Improving Instruction Following},
   author={Renze Lou and Kai Zhang and Jian Xie and Yuxuan Sun and Janice Ahn and Hanzi Xu and Yu su and Wenpeng Yin},
   booktitle={The Twelfth International Conference on Learning Representations},
   year={2024},
   url={https://openreview.net/forum?id=1vrS1zwekw}
}
```