File size: 5,623 Bytes
1ade91a
 
 
c0baf6a
cf11e94
1ade91a
 
 
 
 
 
 
0952a5d
1ade91a
 
 
 
 
 
 
 
 
19bea8a
1ade91a
cbe9672
 
46d96f8
1ade91a
 
9cef06a
 
 
 
 
 
 
 
 
 
 
1ade91a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
cbe9672
 
64e02ae
cbe9672
 
 
 
 
1ade91a
 
 
 
47507f1
1ade91a
f2f108d
 
 
 
 
1ade91a
 
 
 
 
 
 
f05665c
 
 
 
 
 
 
 
 
 
 
 
1ade91a
9cef06a
f05665c
9cef06a
 
 
1ade91a
f05665c
 
 
1ade91a
f05665c
 
104db24
 
 
 
 
 
 
 
 
 
 
 
 
 
0952a5d
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
---
license: mit
datasets:
- croissantllm/croissant_dataset
- croissantllm/CroissantLLM-2201-sft
- cerebras/SlimPajama-627B
- uonlp/CulturaX
- pg19
- bigcode/starcoderdata
language:
- fr
- en
pipeline_tag: text-generation
tags:
- legal
- code
- text-generation-inference
- art
---

# CroissantLLMChat (190k steps + Chat)

This model is part of the CroissantLLM initiative, and corresponds to the checkpoint after 190k steps (2.99 T) tokens and a final Chat finetuning phase.

https://arxiv.org/abs/2402.00786

For best performance, it should be used with a temperature of 0.3 or more, and with the exact template described below:

```python
chat = [
   {"role": "user", "content": "Que puis-je faire à Marseille en hiver?"},
]

chat_input = tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)
```

corresponding to:

```python
chat_input = """<|im_start|>user
{USER QUERY}<|im_end|>
<|im_start|>assistant\n"""
```


## Abstract
We introduce CroissantLLM, a 1.3B language model pretrained on a set of 3T English and French tokens, to bring to the research and industrial community a high-performance, fully open-sourced bilingual model that runs swiftly on consumer-grade local hardware.
To that end, we pioneer the approach of training an intrinsically bilingual model with a 1:1 English-to-French pretraining data ratio, a custom tokenizer, and bilingual finetuning datasets. We release the training dataset, notably containing a French split with manually curated, high-quality, and varied data sources.
To assess performance outside of English, we craft a novel benchmark, FrenchBench, consisting of an array of classification and generation tasks, covering various orthogonal aspects of model performance in the French Language. Additionally, rooted in transparency and to foster further Large Language Model research, we release codebases, and dozens of checkpoints across various model sizes, training data distributions, and training steps, as well as fine-tuned Chat models, and strong translation models. We evaluate our model through the FMTI framework, and validate 81% of the transparency criteria, far beyond the scores of even most open initiatives.
This work enriches the NLP landscape, breaking away from previous English-centric work in order to strengthen our understanding of multilinguality in language models.

## Citation

Our work can be cited as:

```bash
@misc{faysse2024croissantllm,
      title={CroissantLLM: A Truly Bilingual French-English Language Model}, 
      author={Manuel Faysse and Patrick Fernandes and Nuno M. Guerreiro and António Loison and Duarte M. Alves and Caio Corro and Nicolas Boizard and João Alves and Ricardo Rei and Pedro H. Martins and Antoni Bigata Casademunt and François Yvon and André F. T. Martins and Gautier Viaud and Céline Hudelot and Pierre Colombo},
      year={2024},
      eprint={2402.00786},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}
```

## Usage

This model is a Chat model, that is, it is finetuned for Chat function and works best with the provided template.


#### With generate

This might require a stopping criteria on <|im_end|> token.

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


model_name = "croissantllm/CroissantLLMChat-v0.1"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)


generation_args = {
    "max_new_tokens": 256,
    "do_sample": True,
    "temperature": 0.3,
    "top_p": 0.90,
    "top_k": 40,
    "repetition_penalty": 1.05,
    "eos_token_id": [tokenizer.eos_token_id, 32000],
}

chat = [
   {"role": "user", "content": "Qui est le président francais actuel ?"},
]

chat_input = tokenizer.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)

inputs = tokenizer(chat_input, return_tensors="pt").to(model.device)
tokens = model.generate(**inputs, **generation_args)

print(tokenizer.decode(tokens[0]))
# print tokens individually
print([(tokenizer.decode([tok]), tok) for tok in tokens[0].tolist()])
```


## Model limitations

Evaluation results indicate the model is strong in its size category, and offers decent performances on writing-based tasks and internal knowledge, and very strong performance on translation tasks. The small size of the CroissantLLM model however hinders its capacity to perform more complex reasoning-based tasks, at least in a zero or few-shot manner in its generalist base or chat-model versions. This is aligned with other models of size and underlines the importance of scale for more abstract tasks.

#### Knowledge Cutoff 
The model training dataset has a data cutoff date corresponding to the November 2023 Wikipedia dump. This is the de facto knowledge cutoff date for our base model, although a lot of information dates back further. Updated versions can be trained through continued pre-training or subsequent fine-tuning.

#### Multilingual performance.
CroissantLLM is mostly a French and English model. Code performance is relatively limited, and although some amount of data from other languages is included within the SlimPajama training set, out-of-the-box performance in other languages is not to be expected, although some European languages do work quite well. 

#### Hallucinations.
CroissantLLM can hallucinate and output factually incorrect data, especially regarding complex topics. This is to be expected given the small model size, and hallucination rates seem inferior to most models of the same size category although no quantitative assessments have been conducted outside of MT-Bench experiments.