File size: 11,985 Bytes
59290ca
 
 
028d295
59290ca
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a75cdc0
27464b7
 
99b17ec
 
 
 
 
 
 
 
 
 
 
 
 
59290ca
 
fd4a9a8
 
028d295
59290ca
44f3fe5
028d295
44f3fe5
 
028d295
 
59290ca
028d295
59290ca
4d349c9
59290ca
028d295
 
 
44f3fe5
028d295
 
 
8b95958
 
 
 
 
 
028d295
8b95958
 
028d295
8b95958
 
028d295
 
8b95958
 
 
 
 
028d295
8b95958
028d295
8b95958
59290ca
44f3fe5
 
 
 
f0d4457
44f3fe5
 
 
 
59290ca
44f3fe5
59290ca
44f3fe5
59290ca
44f3fe5
59290ca
44f3fe5
 
29e57a5
59290ca
44f3fe5
 
59290ca
44f3fe5
59290ca
44f3fe5
 
 
 
59290ca
 
 
 
44f3fe5
 
 
 
 
 
 
 
 
 
 
 
59290ca
44f3fe5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
59290ca
4cb792b
 
 
59290ca
 
f0d4457
44f3fe5
 
 
 
 
8c53520
 
 
 
 
44f3fe5
 
 
8c53520
ad0c46c
8c53520
 
44f3fe5
 
8c53520
 
 
 
44f3fe5
 
59290ca
bc34f28
59290ca
44f3fe5
59290ca
 
 
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
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
---
datasets:
- bigscience/xP3
license: bigscience-bloom-rail-1.0
language:
- ak
- ar
- as
- bm
- bn
- ca
- code
- en
- es
- eu
- fon
- fr
- gu
- hi
- id
- ig
- ki
- kn
- lg
- ln
- ml
- mr
- ne
- nso
- ny
- or
- pa
- pt
- rn
- rw
- sn
- st
- sw
- ta
- te
- tn
- ts
- tum
- tw
- ur
- vi
- wo
- xh
- yo
- zh
- zu
programming_language: 
- C
- C++
- C#
- Go
- Java
- JavaScript
- Lua
- PHP
- Python
- Ruby
- Rust
- Scala
- TypeScript
pipeline_tag: text-generation
widget:
- text: "一个传奇的开端,一个不灭的神话,这不仅仅是一部电影,而是作为一个走进新时代的标签,永远彪炳史册。Would you rate the previous review as positive, neutral or negative?"
  example_title: "zh-en sentiment"
- text: "一个传奇的开端,一个不灭的神话,这不仅仅是一部电影,而是作为一个走进新时代的标签,永远彪炳史册。你认为这句话的立场是赞扬、中立还是批评?"
  example_title: "zh-zh sentiment"
- text: "Suggest at least five related search terms to \"Mạng neural nhân tạo\"."
  example_title: "vi-en query"
- text: "Proposez au moins cinq mots clés concernant «Réseau de neurones artificiels»."
  example_title: "fr-fr query"
- text: "Explain in a sentence in Telugu what is backpropagation in neural networks."
  example_title: "te-en qa"
- text: "Why is the sky blue?"
  example_title: "en-en qa"
- text: "Write a fairy tale about a troll saving a princess from a dangerous dragon. The fairy tale is a masterpiece that has achieved praise worldwide and its moral is \"Heroes Come in All Shapes and Sizes\". Story (in Spanish):"
  example_title: "es-en fable"
- text: "Write a fable about wood elves living in a forest that is suddenly invaded by ogres. The fable is a masterpiece that has achieved praise worldwide and its moral is \"Violence is the last refuge of the incompetent\". Fable (in Hindi):"
  example_title: "hi-en fable"
---

![xmtf](https://github.com/bigscience-workshop/xmtf/blob/master/xmtf_banner.png?raw=true)

#  Table of Contents

1. [Model Summary](#model-summary)
2. [Use](#use)
3. [Limitations](#limitations)
4. [Training](#training)
5. [Evaluation](#evaluation)
7. [Citation](#citation)

# Model Summary

> We present BLOOMZ & mT0, a family of models capable of following human instructions in dozens of languages zero-shot. We finetune BLOOM & mT5 pretrained multilingual language models on our crosslingual task mixture (xP3) and find our resulting models capable of crosslingual generalization to unseen tasks & languages.

- **Repository:** [bigscience-workshop/xmtf](https://github.com/bigscience-workshop/xmtf)
- **Paper:** [TODO]
- **Point of Contact:** [Niklas Muennighoff](mailto:niklas@hf.co)
- **Languages:** Refer to [BLOOM](https://huggingface.co/bigscience/bloom) for pretraining & [xP3](https://huggingface.co/bigscience/xP3) for finetuning language proportions. It understands both pretraining & finetuning languages.
- **BLOOMZ & mT0 Model Family:**
|Name|Explanation|
|----|-----------|
|[bloomz-560m](https://huggingface.co/bigscience/bloomz-560m)| 560M parameter multitask finetuned version of [bloom-560m](https://huggingface.co/bigscience/bloom-560m) on [xP3](https://huggingface.co/datasets/bigscience/xP3)|
|[bloomz-1b1](https://huggingface.co/bigscience/bloomz-1b1)| 1.1B parameter multitask finetuned version of [bloom-1b1](https://huggingface.co/bigscience/bloom-1b1) on [xP3](https://huggingface.co/datasets/bigscience/xP3)|
|[bloomz-1b7](https://huggingface.co/bigscience/bloomz-1b7)| 1.7B parameter multitask finetuned version of [bloom-1b7](https://huggingface.co/bigscience/bloom-1b7) on [xP3](https://huggingface.co/datasets/bigscience/xP3)|
|[bloomz-3b](https://huggingface.co/bigscience/bloomz-3b)| 3B parameter multitask finetuned version of [bloom-3b](https://huggingface.co/bigscience/bloom-3b) on [xP3](https://huggingface.co/datasets/bigscience/xP3)|
|[bloomz-7b1](https://huggingface.co/bigscience/bloomz-7b1)|7.1B parameter multitask finetuned version of [bloom-7b1](https://huggingface.co/bigscience/bloom-7b1) on [xP3](https://huggingface.co/datasets/bigscience/xP3)|
|[bloomz](https://huggingface.co/bigscience/bloomz)|176B parameter multitask finetuned version of [bloom](https://huggingface.co/bigscience/bloom) on [xP3](https://huggingface.co/datasets/bigscience/xP3)|
|||
|[bloomz-7b1-mt](https://huggingface.co/bigscience/bloomz-7b1-mt)|7.1B parameter multitask finetuned version of [bloom-7b1](https://huggingface.co/bigscience/bloom-7b1) on [xP3](https://huggingface.co/datasets/bigscience/xP3) & [xP3mt](https://huggingface.co/bigscience/datasets/xP3mt). **Better than [bloomz-7b1](https://huggingface.co/bigscience/bloomz-7b1) when prompting in non-English**|
|[bloomz-mt](https://huggingface.co/bigscience/bloomz-mt)| 176B parameter multitask finetuned version of [bloom](https://huggingface.co/bigscience/bloom) on [xP3](https://huggingface.co/datasets/bigscience/xP3) & [xP3mt](https://huggingface.co/bigscience/datasets/xP3mt). **Better than [bloomz](https://huggingface.co/bigscience/bloomz) when prompting in non-English**|
|||
|[bloomz-7b1-p3](https://huggingface.co/bigscience/bloomz-7b1)| 7.1B parameter multitask finetuned version of [bloom-7b1](https://huggingface.co/bigscience/bloom-7b1) on [P3](https://huggingface.co/datasets/bigscience/P3). **Released for research purposes, performance is inferior to [bloomz-7b1](https://huggingface.co/bigscience/bloomz-7b1)**|
|[bloomz-p3](https://huggingface.co/bigscience/bloomz)| 176B parameter multitask finetuned version of [bloom](https://huggingface.co/bigscience/bloom) on [P3](https://huggingface.co/datasets/bigscience/P3). **Released for research purposes, performance is inferior to [bloomz](https://huggingface.co/bigscience/bloomz)**|
|||
|||
|[mt0-small](https://huggingface.co/bigscience/mt0-xxl)|300M parameter multitask finetuned version of [mt5-small](https://huggingface.co/google/mt5-small) on [xP3](https://huggingface.co/datasets/bigscience/xP3)|
|[mt0-base](https://huggingface.co/bigscience/mt0-xxl)|580M parameter multitask finetuned version of [mt5-base](https://huggingface.co/google/mt5-base) on [xP3](https://huggingface.co/datasets/bigscience/xP3)|
|[mt0-large](https://huggingface.co/bigscience/mt0-xxl)|1.2B parameter multitask finetuned version of [mt5-large](https://huggingface.co/google/mt5-large) on [xP3](https://huggingface.co/datasets/bigscience/xP3)|
|[mt0-xl](https://huggingface.co/bigscience/mt0-xxl)|3.7B parameter multitask finetuned version of [mt5-xl](https://huggingface.co/google/mt5-xl) on [xP3](https://huggingface.co/datasets/bigscience/xP3)|
|[mt0-xxl](https://huggingface.co/bigscience/mt0-xxl)|13B parameter multitask finetuned version of [mt5-xxl](https://huggingface.co/google/mt5-xxl) on [xP3](https://huggingface.co/datasets/bigscience/xP3)|
|||
|[mt0-xxl-mt](https://huggingface.co/bigscience/mt0-xxl-mt)|13B parameter multitask finetuned version of [mt5-xxl](https://huggingface.co/google/mt5-xxl) on [xP3](https://huggingface.co/datasets/bigscience/xP3) & [xP3mt](https://huggingface.co/datasets/bigscience/xP3mt). **Better than [mt0-xxl](https://huggingface.co/bigscience/mt0-xxl) when prompting in non-English**|
|||
|[mt0-xxl-p3](https://huggingface.co/bigscience/mt0-xxl-p3)| 13B parameter multitask finetuned version of [mt5-xxl](https://huggingface.co/google/mt5-xxl) on [P3](https://huggingface.co/datasets/bigscience/P3). **Released for research purposes, performance is inferior to [mt0-xxl](https://huggingface.co/bigscience/mt0-xxl)**|

# Use

## Intended use

We recommend using the model to perform tasks expressed in natural language. For example, given the prompt "*Translate to English: Je t’aime.*", the model will most likely answer "*I love you.*". Some prompt ideas from our paper: 
- 一个传奇的开端,一个不灭的神话,这不仅仅是一部电影,而是作为一个走进新时代的标签,永远彪炳史册。你认为这句话的立场是赞扬、中立还是批评?
- Suggest at least five related search terms to "Mạng neural nhân tạo".
- Write a fairy tale about a troll saving a princess from a dangerous dragon. The fairy tale is a masterpiece that has achieved praise worldwide and its moral is "Heroes Come in All Shapes and Sizes". Story (in Spanish):
- Explain in a sentence in Telugu what is backpropagation in neural networks.

**Feel free to share your generations in the Community tab!**

## How to use

### CPU

<details>
<summary> Click to expand </summary>

```python
# pip install -q transformers
from transformers import AutoModelForCausalLM, AutoTokenizer

checkpoint = "bigscience/bloomz"

tokenizer = AutoTokenizer.from_pretrained(checkpoint)
model = AutoModelForCausalLM.from_pretrained(checkpoint)

inputs = tokenizer.encode("Translate to English: Je t’aime.", return_tensors="pt")
outputs = model.generate(inputs)
print(tokenizer.decode(outputs[0]))
```

</details>

### GPU

<details>
<summary> Click to expand </summary>

```python
# pip install -q transformers accelerate
from transformers import AutoModelForCausalLM, AutoTokenizer

checkpoint = "bigscience/bloomz"

tokenizer = AutoTokenizer.from_pretrained(checkpoint)
model = AutoModelForCausalLM.from_pretrained(checkpoint, torch_dtype="auto", device_map="auto")

inputs = tokenizer.encode("Translate to English: Je t’aime.", return_tensors="pt").to("cuda")
outputs = model.generate(inputs)
print(tokenizer.decode(outputs[0]))
```

</details>

### GPU in 8bit

<details>
<summary> Click to expand </summary>

```python
# pip install -q transformers accelerate bitsandbytes
from transformers import AutoModelForCausalLM, AutoTokenizer

checkpoint = "bigscience/bloomz"

tokenizer = AutoTokenizer.from_pretrained(checkpoint)
model = AutoModelForCausalLM.from_pretrained(checkpoint, device_map="auto", load_in_8bit=True)

inputs = tokenizer.encode("Translate to English: Je t’aime.", return_tensors="pt").to("cuda")
outputs = model.generate(inputs)
print(tokenizer.decode(outputs[0]))
```

</details>

<!-- Necessary for whitespace -->
###

# Limitations

**Prompt Engineering:** The performance may vary depending on the prompt. For BLOOMZ models, we recommend making it very clear when the input stops to avoid the model trying to continue it. For example, the prompt "*Translate to English: Je t'aime*" without the full stop (.) at the end, may result in the model trying to continue the French sentence. Better prompts are e.g. "*Translate to English: Je t'aime.*", "*Translate to English: Je t'aime. Translation:*" "*What is "Je t'aime." in English?*", where it is clear for the model when it should answer. Further, we recommend providing the model as much context as possible. For example, if you want it to answer in Telugu, then tell the model, e.g. "*Explain in a sentence in Telugu what is backpropagation in neural networks.*".

# Training

## Model

- **Architecture:** Same as [bloom](https://huggingface.co/bigscience/bloom), also refer to the `config.json` file
- **Finetuning steps:** 498
- **Finetuning tokens:** 2.09 billion
- **Finetuning layout:** 72x pipeline parallel, 1x tensor parallel, 4x data parallel
- **Precision:** bfloat16

## Hardware

- **CPUs:** AMD CPUs with 512GB memory per node
- **GPUs:** 288 A100 80GB GPUs with 8 GPUs per node (36 nodes) using NVLink 4 inter-gpu connects, 4 OmniPath links
- **Communication:** NCCL-communications network with a fully dedicated subnet

## Software

- **Orchestration:** [Megatron-DeepSpeed](https://github.com/bigscience-workshop/Megatron-DeepSpeed)
- **Optimizer & parallelism:** [DeepSpeed](https://github.com/microsoft/DeepSpeed)
- **Neural networks:** [PyTorch](https://github.com/pytorch/pytorch) (pytorch-1.11 w/ CUDA-11.5)
- **FP16 if applicable:** [apex](https://github.com/NVIDIA/apex)

# Evaluation

We refer to Table 7 from our paper [TODO LINK] and [bigscience/evaluation-results](https://huggingface.co/datasets/bigscience/evaluation-results).

# Citation
```bibtex
TODO
```