File size: 1,705 Bytes
e99ff02
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41

---
language:
- ru
tags:
- generation
- question answering
- instruction tuning
license: cc-by-nc-4.0
---

###  Model Description

This HF repository contains base LLMs instruction tuned (SFT) with LoRA and then used to study whether monolingual or multilingual instruction tuning is more favourable.
* [GitHub](https://github.com/hplt-project/monolingual-multilingual-instruction-tuning/tree/main)
* [Paper](https://arxiv.org/abs/2309.08958)

#### Instruction tuning details
* Base model: [bigscience/bloom-7b1](https://huggingface.co/bigscience/bloom-7b1)
* Instruction tuning language: Russian
* Training method: LoRA.
* LoRA details: rank=8, alpha=16, target modules={key, query, value}.
* Best checkpoint: best cross-entropy on a validation set, trained for 5 epochs.
* Dataset: machine-translated from [yahma/alpaca-cleaned](https://huggingface.co/datasets/yahma/alpaca-cleaned). You can download our data [HERE](https://github.com/hplt-project/monolingual-multilingual-instruction-tuning/tree/main/training-data).

#### Usage
The model checkpoint should be loaded with the base model together using `transformers` and `peft` libraries.

Please refer to our Github repository [HERE](https://github.com/hplt-project/monolingual-multilingual-instruction-tuning/tree/main/loraft) for inference and training instructions.

#### Citation
```
@inproceedings{chen-etal-2024-monolingual,
  title="Monolingual or multilingual instruction tuning: Which makes a better {Alpaca}",
  author="Pinzhen Chen and Shaoxiong Ji and Nikolay Bogoychev and Andrey Kutuzov and Barry Haddow and Kenneth Heafield",
  year="2024",
  booktitle = "Findings of the Association for Computational Linguistics: EACL 2024",
}
```