File size: 16,765 Bytes
46cab60
aca8cfd
 
 
 
 
 
 
 
 
bede6e8
aca8cfd
 
 
 
 
46cab60
aca8cfd
 
 
 
 
8eceaae
aca8cfd
c8f621c
aca8cfd
 
 
c8f621c
aca8cfd
 
 
 
 
 
 
 
 
 
 
 
 
8eceaae
aca8cfd
8eceaae
aca8cfd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b3a294d
aca8cfd
 
 
 
 
 
 
 
 
69e469f
aca8cfd
 
69e469f
aca8cfd
 
 
 
72e9a69
aca8cfd
 
 
e5ee700
aca8cfd
 
d0a30cf
aca8cfd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
69e469f
 
aca8cfd
 
 
69e469f
 
 
 
d0a30cf
69e469f
 
e5ee700
69e469f
 
d0a30cf
69e469f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e5ee700
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
69e469f
aca8cfd
 
 
 
 
e5ee700
aca8cfd
 
69e469f
aca8cfd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
69e469f
 
 
aca8cfd
 
 
 
 
 
 
 
 
 
 
 
 
 
69e469f
 
 
aca8cfd
 
 
 
 
 
 
 
 
44b3209
 
 
 
 
 
 
69e469f
 
aca8cfd
 
 
 
 
 
 
 
 
 
 
 
 
69e469f
 
aca8cfd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
69e469f
 
 
aca8cfd
 
 
 
 
 
 
 
 
 
 
 
 
 
69e469f
 
aca8cfd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
36f2b77
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
aca8cfd
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
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
---
language:
- 'no'
- nb
- nn
inference: true
tags:
- mistral
- gpt
- generative
license: apache-2.0
pipeline_tag: text-generation
datasets:
- uonlp/CulturaX
- NbAiLab/NCC
- vikp/starcoder_filtered
---

# **NorMistral-7b-scratch**

<img align="center" src="https://huggingface.co/ltg/norbert3-base/resolve/main/norbert.png" width=12.5%>

NorMistral-7b-scratch is a large Norwegian language model pretrained from scratch on a total of 260 billion subword tokens (using six repetitions of open Norwegian texts).

This model is a part of the NORA.LLM family developed in collaboration between [the Language Technology Group at the University of Oslo](https://huggingface.co/ltg), [the High Performance Language Technologies (HPLT) project](https://hplt-project.org/), [the National Library of Norway](https://huggingface.co/NbAiLab), and [the University of Turku](https://huggingface.co/TurkuNLP).
All the models are pre-trained on the same dataset and with the same tokenizer.
NorMistral-7b-scratch has over 7 billion parameters and is based on [the Mistral architecture](https://huggingface.co/mistralai/Mistral-7B-v0.1).

The NORA.LLM language model family includes (as of now):
- [**NorMistral-7b-warm**](https://huggingface.co/norallm/normistral-7b-warm) -- an LLM initialized from [Mistral-7b-v0.1](https://huggingface.co/mistralai/Mistral-7B-v0.1) and continuously pretrained on Norwegian data;
- [**NorMistral-7b-scratch**](https://huggingface.co/norallm/normistral-7b-scratch) -- a Mistral-based LLM pretrained from scratch on Norwegian data;
- [**NorBLOOM-7b-scratch**](https://huggingface.co/norallm/NorBLOOM-7b-scratch) -- a BLOOM-based LLM pretrained from scratch on Norwegian data.


*Disclaimer: This model is pretrained on raw (mostly web-based) textual data.
It is not finetuned to follow instructions, and it can generate harmful completions after inappropriate user prompts.
It is primarily intended for research purposes.*

_____
## Pretraining corpus

The model is pretrained exclusively on publicly available data. We combine the resources from [the public part of the NCC corpus](https://huggingface.co/datasets/NbAiLab/NCC), from [the cleaned HPLT corpus](https://hplt-project.org/datasets/v1.2), and from [CulturaX](https://huggingface.co/datasets/uonlp/CulturaX).
This resulted in over 34B subword tokens of Norwegian (Bokmål or Nynorsk) in total, which amounts to about 26.7B whitespace-separated tokens.
We also augment the corpus with [Starcoder](https://huggingface.co/datasets/vikp/starcoder_filtered); 20% of the 260B tokens are sampled from this code corpus.
The natural language data is repeated six times to get the pretraining budget of 260B tokens, in accordance with findings from [Muennighoff et al. (2023)](https://neurips.cc/virtual/2023/poster/70706).


_____
## Model details

**Model Developers:** Language Technology Group at the University of Oslo.

**Variations:** NorMistral is currently published as two 7B variants: one trained entirely from *scratch* and one *warm*-started from the Mistral model.

**Input:** Textual input.

**Output:** Generated text.

**Model Architecture:** NorMistral is an auto-regressive language model that uses an optimized transformer architecture based on the Mistral/Llama language models.

||Training Data|Params|Context Length|Tokens|LR|
|---|---|---|---|---|---|
|NorMistral-7b-warm|NCC+HPLT+CulturaX+Starcoder|7B|2k|260B|1.0 x 10<sup>-4</sup>|
|NorMistral-7b-scratch|NCC+HPLT+CulturaX+Starcoder|7B|2k|260B|3.0 x 10<sup>-4</sup>|
|NorBLOOM-7b-scratch|NCC+HPLT+CulturaX+Starcoder|7B|2k|260B|1.2 x 10<sup>-4</sup>|

**Tokenizer:** Byte-based BPE tokenizer trained on the same Norwegian corpus as this model. The vocabulary size is 32,768 tokens.

**Training FLOPs** The approximate amount is 1.22e+22 FLOPs; calculated as in [Chowdhery et al. (2022)](https://arxiv.org/abs/2204.02311).

**Model Dates:** The models were pretrained between December 2023 and January 2024.

**Status:** These are only pretrained language models; instruction-finetuned models will follow soon.

**License:** [Apache-2.0](https://www.apache.org/licenses/LICENSE-2.0)

**Research Paper:** Forthcoming

_____
## Initial evaluation

*Disclaimer: our model evaluation is an ongoing phase and is not claimed to be exhaustive. We provide our initial evaluation results on standard natural language understanding and generation tasks, and our evaluation design will be extended. 
The user should perform evaluation for their particular model application scenario, including safety and bias evaluations.*

The perplexity on the heldout [validation set from the Norwegian Colossal Corpus (NCC)](https://huggingface.co/datasets/NbAiLab/NCC) is 7.43 and the final training perplexity is 4.76.

Our initial downstream evaluation is conducted on reading comprehension, sentiment analysis and machine translation tasks using open-source peer-reviewed datasets and benchmarks in native Norwegian.
We release [our codebase here](https://github.com/ltgoslo/norallm). We compare against other pretrained generative language models that officially support Norwegian: [NB-GPT-J](https://huggingface.co/NbAiLab/nb-gpt-j-6B), [GPT-Sw3 6.7B](https://huggingface.co/AI-Sweden-Models/gpt-sw3-6.7b), [GPT-Sw3 6.7B v2](https://huggingface.co/AI-Sweden-Models/gpt-sw3-6.7b-v2), and [Falcon-7B](https://huggingface.co/tiiuae/falcon-7b); we also include evaluation of [Mistral-7b-v0.1](https://huggingface.co/mistralai/Mistral-7B-v0.1).


### Sentiment analysis

[NoReC](https://huggingface.co/datasets/ltg/norec_sentence) ([Øvrelid et al., 2020](https://aclanthology.org/2020.lrec-1.618/)) is a dataset for sentence-level sentiment analysis derived from the Norwegian Review Corpus [(Velldal et al., 2018)](https://aclanthology.org/L18-1661/).
We use the binary formulation of this task (positive vs. negative).

<details>
<summary>Method (click to expand)</summary>
  
* Evaluation setting: zero-shot and few-shot perplexity-based evaluation.
* Prompt: ```"Tekst: {text}\nSentiment:{label}"```, where the ```label``` is either "positiv" or "negativ".
* Few-shot results show the average scores across 5 repetitions
* Evaluation script: https://github.com/ltgoslo/norallm/blob/main/initial_evaluation/sentiment_analysis.py
* Performance metric: macro-averaged F1-score.

</details>

<details open>
<summary>Macro-averaged F1-scores on the sentence-level sentiment analysis task (NoReC)</summary>

|Model|0-shot (macro F1)|1-shot (macro F1)|16-shot (macro F1)|
|---|---|---|---|
|NorMistral-7b-warm|60.6|**77.8**|**87.3**|
|NorMistral-7b-scratch|47.3|62.2|80.1|
|NorBLOOM-7b|**75.7**|73.8|65.5|
|NB-GPT-J|48.4|56.5|65.2|
|GPT-Sw3-6.7B|61.5|72.2|76.5|
|GPT-Sw3-6.7B-v2|42.4|69.1|83.4|
|Falcon-7B|53.3|61.6|74.9|
|Mistral-7B-v0.1|70.2|72.9|84.8|

</details>



### Reading comprehension

[NorQuAD](https://huggingface.co/datasets/ltg/norquad) ([Ivanova et al., 2023](https://aclanthology.org/2023.nodalida-1.17/)) is a dataset for extractive question answering in Norwegian designed similarly to [SQuAD (Rajpurkar et al., 2016)](https://aclanthology.org/D16-1264/). 

<details>
<summary>Method (click to expand)</summary>
  
* Evaluation setting: zero-shot and few-shot settings via natural language generation using the greedy decoding strategy.
* Prompt: ```"Tittel: {title}\n\nTekst: {text}\n\nSpørsmål: {question}\n\nSvar:{answer}"``` Based on [Brown et al. (2020)](https://arxiv.org/abs/2005.14165).
* Few-shot results show the average scores across 5 repetitions
* Evaluation script: https://github.com/ltgoslo/norallm/blob/main/initial_evaluation/norquad.py
* Performance metrics: macro-averaged F1-score and exact match (EM).
  
</details>

<details open>
<summary>Performance results on the extractive question answering task (NorQuAD)</summary>

|Model|0-shot (F1/EM)|1-shot (F1/EM)|2-shot (F1/EM)|
|---|---|---|---|
|NorMistral-7b-warm|**48.6**/**24.8**|63.6/40.0|66.5/43.8|
|NorMistral-7b-scratch|34.0/15.7|46.5/25.8|48.5/27.8|
|NorBLOOM-7b|35.0/13.3|47.7/28.0|49.3/30.1|
|NB-GPT-J|24.4/6.8|32.8/11.6|35.0/12.3|
|GPT-Sw3-6.7B|46.5/22.0|55.9/32.0|58.1/34.3|
|GPT-Sw3-6.7B-v2|46.9/22.5|61.1/38.9|66.0/44.5|
|Falcon-7B|15.8/7.0|27.3/13.9|27.4/13.1|
|Mistral-7B-v0.1|46.4/22.4|**64.9**/**41.1**|**71.7**/**49.4**|

</details>


### Grammatical error correction

[ASK-RAW](https://huggingface.co/datasets/ltg/ask-gec) is dataset for Norwegian grammatical error correction (GEC) created by [Matias Jentoft (2023)](https://www.duo.uio.no/handle/10852/103885).

<details>
<summary>Method (click to expand)</summary>
  
* Evaluation setting: zero-shot and few-shot settings via natural language generation using the greedy decoding strategy.
* Prompt: ```"Her er eksempler på perfekt korrigering av grammatiske feil:\n\nTekst: {source_text}\nKorreksjon:{target_text}"```
* Few-shot results show the average scores across 5 repetitions
* Evaluation script: https://github.com/ltgoslo/norallm/blob/main/initial_evaluation/gec.py
* Performance metrics: the evaluation metric uses [ERRANT](https://github.com/chrisjbryant/errant/tree/main), which identifies edit-spans and then calculates the F_{0.5} scores between the gold edits and predicted edits. 
  
</details>

<details open>
<summary>Results on [the ASK corpus](https://huggingface.co/datasets/ltg/ask-gec) (ERRANT F_{0.5})</summary>

|Model|0-shot (F0.5)|1-shot (F0.5)|32-shot (F0.5)|
|---|---|---|---|
|NorMistral-7b-warm|**40.8**|41.8|48.5|
|NorMistral-7b-scratch|22.1|28.8|42.1|
|NorBLOOM-7b|8.7|24.5|32.0|
|NB-GPT-J|9.1|28.2|30.6|
|GPT-Sw3-6.7B|30.5|42.9|**50.6**|
|GPT-Sw3-6.7B-v2|40.6|**43.4**|49.8|
|Falcon-7B|10.8|12.4|15.5|
|Mistral-7B-v0.1|26.0|27.4|30.6|

</details>


### Machine translation

[Tatoeba](https://huggingface.co/datasets/Helsinki-NLP/tatoeba_mt) [(Tiedemann, 2020)](https://aclanthology.org/2020.wmt-1.139/) is a benchmark for machine translation, which includes hundreds of language pairs. We consider six language pairs (English <-> Bokmål, English <-> Nynorsk, and Bokmål <-> Nynorsk).

<details>
<summary>Method (click to expand)</summary>
  
* Evaluation setting: zero-shot and few-shot settings via natural language generation using the greedy decoding strategy.
* Prompt: ```"{source_language}: {source_text}\n{target_language}:{target_text}"```, where the ```source_language``` and ```target_language``` are ```Engelsk```, ```Bokmål```, or ```Nynorsk```. Based on [Garcia et al. (2023)](https://arxiv.org/abs/2302.01398).
* Few-shot results show the average scores across 5 repetitions
* Evaluation script: https://github.com/ltgoslo/norallm/blob/main/initial_evaluation/machine_translation.py
* Performance metrics: BLEU ([Papineni et al., 2002](https://aclanthology.org/P02-1040/)) and chrF++ ([Popović, 2015](https://aclanthology.org/W15-3049/)).

</details>

<details open>
<summary>English → Norwegian Bokmål</summary>

|Model|0-shot (BLEU/chrF++)|1-shot (BLEU/chrF++)|5-shot (BLEU/chrF++)|
|---|---|---|---|
|NorMistral-7b-warm|**55.8**/**70.7**|**56.7**/**71.5**|57.7/72.4|
|NorMistral-7b-scratch|46.4/62.9|50.4/66.3|52.1/67.6|
|NorBLOOM-7b|37.1/53.6|50.1/65.8|52.0/67.6|
|NB-GPT-J|8.6/39.1|35.9/64.5|47.2/68.7|
|GPT-Sw3-6.7B|21.8/55.2|54.5/69.6|**58.6**/**73.2**|
|GPT-Sw3-6.7B-v2|20.6/53.2|51.2/66.6|58.4/73.0|
|Falcon-7B|19.1/40.1|20.6/41.8|22.1/43.6|
|Mistral-7B-v0.1|32.5/51.9|35.4/55.1|36.3/56.0|


</details>

<details open>
<summary>English → Norwegian Nynorsk</summary>

|Model|0-shot (BLEU/chrF++)|1-shot (BLEU/chrF++)|5-shot (BLEU/chrF++)|
|---|---|---|---|
|NorMistral-7b-warm|**43.6**/**62.0**|**44.2**/**63.2**|44.3/**63.7**|
|NorMistral-7b-scratch|38.0/56.9|39.2/57.9|40.7/59.3|
|NorBLOOM-7b|35.6/54.7|36.6/56.3|38.1/57.4|
|NB-GPT-J|1.7/14.7|6.3/34.1|35.2/60.4|
|GPT-Sw3-6.7B|13.4/44.3|43.6/62.5|**44.5**/63.5|
|GPT-Sw3-6.7B-v2|14.8/45.5|43.7/62.3|44.0/63.6|
|Falcon-7B|6.4/28.6|8.3/30.5|9.3/32.1|
|Mistral-7B-v0.1|11.6/35.7|13.5/38.7|15.0/40.0|


</details>


<details open>
<summary>Norwegian Bokmål → English</summary>

|Model|0-shot (BLEU/chrF++)|1-shot (BLEU/chrF++)|5-shot (BLEU/chrF++)|
|---|---|---|---|
|NorMistral-7b-warm|**56.7**/**70.6**|**57.7**/**71.7**|**58.5**/**72.2**|
|NorMistral-7b-scratch|48.1/62.9|51.5/66.6|52.6/67.6|
|NorBLOOM-7b|46.0/61.5|51.3/66.7|51.7/66.9|
|NB-GPT-J|23.9/55.3|32.3/63.1|48.5/68.7|
|GPT-Sw3-6.7B|47.9/67.8|52.4/70.6|50.0/70.7|
|GPT-Sw3-6.7B-v2|38.8/59.6|49.0/68.6|50.7/70.6|
|Falcon-7B|42.4/58.5|47.3/62.3|48.6/63.3|
|Mistral-7B-v0.1|53.8/68.2|54.6/69.0|56.9/70.7|

</details>

<details open>
<summary>Norwegian Nynorsk → English</summary>

|Model|0-shot (BLEU/chrF++)|1-shot (BLEU/chrF++)|5-shot (BLEU/chrF++)|
|---|---|---|---|
|NorMistral-7b-warm|**55.1**/**68.4**|**55.5**/**69.5**|56.0/69.8|
|NorMistral-7b-scratch|47.1/61.9|49.4/64.2|52.3/66.2|
|NorBLOOM-7b|45.0/59.3|48.3/64.0|49.0/64.7|
|NB-GPT-J|2.9/19.5|10.1/41.0|44.4/66.9|
|GPT-Sw3-6.7B|47.8/66.2|49.1/68.1|49.6/69.4|
|GPT-Sw3-6.7B-v2|46.3/67.5|48.9/69.3|**58.2**/**72.8**|
|Falcon-7B|21.6/40.6|31.7/47.4|36.6/57.1|
|Mistral-7B-v0.1|40.7/57.1|46.2/60.7|49.9/63.8|

</details>


<details open>
<summary>Norwegian Bokmål → Norwegian Nynorsk</summary>

|Model|0-shot (BLEU/chrF++)|1-shot (BLEU/chrF++)|5-shot (BLEU/chrF++)|
|---|---|---|---|
|NorMistral-7b-warm|**75.8**/**87.5**|74.0/**86.9**|75.3/87.5|
|NorMistral-7b-scratch|38.0/56.9|39.2/57.9|40.7/59.3|
|NorBLOOM-7b|71.5/84.4|70.1/84.1|71.9/85.1|
|NB-GPT-J|6.6/35.5|9.6/41.0|26.0/64.7|
|GPT-Sw3-6.7B|63.6/82.8|74.7/86.0|75.8/86.9|
|GPT-Sw3-6.7B-v2|57.5/81.1|**75.3**/86.7|**76.7**/**87.6**|
|Falcon-7B|28.7/59.2|29.8/60.8|32.1/62.3|
|Mistral-7B-v0.1|32.0/62.2|32.9/62.6|35.2/63.9|


</details>

<details open>
<summary>Norwegian Nynorsk → Norwegian Bokmål</summary>

|Model|0-shot (BLEU/chrF++)|1-shot (BLEU/chrF++)|5-shot (BLEU/chrF++)|
|---|---|---|---|
|NorMistral-7b-warm|**88.1**/**93.6**|**89.2**/**94.3**|**89.3**/**94.6**|
|NorMistral-7b-scratch|85.1/91.4|86.6/92.4|87.4/93.0|
|NorBLOOM-7b|78.7/88.5|84.2/90.7|87.4/93.0|
|NB-GPT-J|2.7/18.5|6.9/35.6|52.9/84.3|
|GPT-Sw3-6.7B|652.3/82.4|86.1/92.5|87.8/93.6|
|GPT-Sw3-6.7B-v2|72.0/88.6|86.1/92.5|88.2/93.9|
|Falcon-7B|36.7/61.6|38.3/63.5|45.8/68.1|
|Mistral-7B-v0.1|57.0/74.8|59.9/77.5|62.6/79.1|

</details>


_____
## Hardware and Software

**Training Factors:** The models were pretrained using the Megatron-DeepSpeed library on [the LUMI cluster in Finland](https://lumi-supercomputer.eu/).

**Carbon Footprint:** Pretraining one model took approximately 70k GPU hours of computation on AMD MI250X GPUs (assuming 2 GPUs per one AMD MI250X device), each of which draws 500W.
LUMI is [one of the most eco-efficient data centers in the world](https://www.lumi-supercomputer.eu/sustainable-future/), and its energy consumption is covered 100% with renewable electricity.



_____
## Example usage

Let's try to use this model for English-to-Norwegian machine translation using simple zero-shot prompting:

```python
from transformers import AutoTokenizer, AutoModelForCausalLM

# First, we will have to import the tokenizer and the language model
tokenizer = AutoTokenizer.from_pretrained("norallm/normistral-7b-scratch")
model = AutoModelForCausalLM.from_pretrained("norallm/normistral-7b-scratch").cuda().eval()

# Now we will define the zero-shot prompt template
prompt = """Engelsk: {0}
Bokmål:"""

# A function that will take care of generating the output
@torch.no_grad()
def generate(text):
    text = prompt.format(text)
    input_ids = tokenizer(text, return_tensors='pt').input_ids.cuda()
    prediction = model.generate(
        input_ids,
        max_new_tokens=64,
        do_sample=False,
        eos_token_id=tokenizer('\n').input_ids
    )
    return tokenizer.decode(prediction[0, input_ids.size(1):]).strip()

# Now you can simply call the generate function with an English text you want to translate:
generate("I'm super excited about this Norwegian NORA model! Can it translate these sentences?")
# > this should output: 'Jeg er super spent på denne norske NORA modellen! Kan den oversette disse setningene?'
```

## Example usage on a GPU with ~16GB VRAM (try for yourself [in Google Colab](https://colab.research.google.com/drive/1AQgJ8lN-SNOqkUKj4xpQI5rr0R7V2Xzy?usp=sharing))
Install bitsandbytes if you want to load in 8bit

```bash
pip install bitsandbytes
pip install accelerate
```


```python
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained(
    "norallm/normistral-7b-scratch"
)

# This setup needs about 8gb VRAM
# Setting `load_in_8bit=False` -> 15gb VRAM
# Using `torch.float32` and `load_in_8bit=False` -> 21gb VRAM
model = AutoModelForCausalLM.from_pretrained(
    "norallm/normistral-7b-scratch",
    device_map='auto',
    load_in_8bit=True,
    torch_dtype=torch.bfloat16
)
```