---
license: mit
---

This repo contains a low-rank adapter (LoRA) for BLOOM-7b1 
fit on the [Stanford-Alpaca-52k](https://github.com/tatsu-lab/stanford_alpaca)
and [databricks-dolly-15k](https://github.com/databrickslabs/dolly/tree/master/data) data in English.

### Dataset Creation

1. English Instructions: The English instuctions are obtained from [alpaca-52k](https://github.com/tatsu-lab/stanford_alpaca), and [dolly-15k](https://github.com/databrickslabs/dolly/tree/master/data).
2. Instruction Translation: The instructions (and inputs) are translated into the target languages using Google Translation API (conducted on April 2023).
3. Output Generation: We generate output from `gpt-3.5-turbo` for each language (conducted on April 2023).

<h3 align="center">
<img src="https://raw.githubusercontent.com/fajri91/eval_picts/master/BactrianX_dataset.jpg" width="950" align="center">
</h3>

### Training Parameters

The code for training the model is provided in our [github](https://github.com/mbzuai-nlp/Bactrian-X), which is adapted from [Alpaca-LoRA](https://github.com/tloen/alpaca-lora).
This version of the weights was trained with the following hyperparameters:

- Epochs: 8
- Batch size: 128
- Cutoff length: 1024
- Learning rate: 3e-4
- Lora _r_: 16
- Lora target modules: query_key_value


That is:

```
python finetune.py \
    --base_model='bigscience/bloom-7b1' \
    --num_epochs=5 \
    --cutoff_len=1024 \
    --group_by_length \
    --output_dir='./bactrian-en-bloom-7b1-lora' \
    --lora_target_modules='query_key_value' \
    --lora_r=16 \
    --micro_batch_size=32
```

Instructions for running it can be found at https://github.com/MBZUAI-nlp/Bactrian-X.

### Discussion of Biases

(1) Translation bias; (2) Potential English-culture bias in the translated dataset.


### Citation Information

```
@misc{li2023bactrianx,
      title={Bactrian-X : A Multilingual Replicable Instruction-Following Model with Low-Rank Adaptation}, 
      author={Haonan Li and Fajri Koto and Minghao Wu and Alham Fikri Aji and Timothy Baldwin},
      year={2023},
      eprint={2305.15011},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

```