Triangle104 commited on
Commit
a15d5ef
1 Parent(s): cb99a6c

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +93 -0
README.md CHANGED
@@ -46,6 +46,99 @@ tags:
46
  This model was converted to GGUF format from [`utter-project/EuroLLM-1.7B-Instruct`](https://huggingface.co/utter-project/EuroLLM-1.7B-Instruct) using llama.cpp via the ggml.ai's [GGUF-my-repo](https://huggingface.co/spaces/ggml-org/gguf-my-repo) space.
47
  Refer to the [original model card](https://huggingface.co/utter-project/EuroLLM-1.7B-Instruct) for more details on the model.
48
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
49
  ## Use with llama.cpp
50
  Install llama.cpp through brew (works on Mac and Linux)
51
 
 
46
  This model was converted to GGUF format from [`utter-project/EuroLLM-1.7B-Instruct`](https://huggingface.co/utter-project/EuroLLM-1.7B-Instruct) using llama.cpp via the ggml.ai's [GGUF-my-repo](https://huggingface.co/spaces/ggml-org/gguf-my-repo) space.
47
  Refer to the [original model card](https://huggingface.co/utter-project/EuroLLM-1.7B-Instruct) for more details on the model.
48
 
49
+ ---
50
+ Model details:
51
+ -
52
+ This is the model card for the first instruction tuned model of the EuroLLM series: EuroLLM-1.7B-Instruct. You can also check the pre-trained version: EuroLLM-1.7B.
53
+
54
+ Developed by: Unbabel, Instituto Superior Técnico, University of Edinburgh, Aveni, University of Paris-Saclay, University of Amsterdam, Naver Labs, Sorbonne Université.
55
+ Funded by: European Union.
56
+ Model type: A 1.7B parameter instruction tuned multilingual transfomer LLM.
57
+ Language(s) (NLP): Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Irish, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Slovak, Slovenian, Spanish, Swedish, Arabic, Catalan, Chinese, Galician, Hindi, Japanese, Korean, Norwegian, Russian, Turkish, and Ukrainian.
58
+ License: Apache License 2.0.
59
+
60
+ Model Details
61
+
62
+ The EuroLLM project has the goal of creating a suite of LLMs capable of understanding and generating text in all European Union languages as well as some additional relevant languages. EuroLLM-1.7B is a 1.7B parameter model trained on 4 trillion tokens divided across the considered languages and several data sources: Web data, parallel data (en-xx and xx-en), and high-quality datasets. EuroLLM-1.7B-Instruct was further instruction tuned on EuroBlocks, an instruction tuning dataset with focus on general instruction-following and machine translation.
63
+ Model Description
64
+
65
+ EuroLLM uses a standard, dense Transformer architecture:
66
+
67
+ We use grouped query attention (GQA) with 8 key-value heads, since it has been shown to increase speed at inference time while maintaining downstream performance.
68
+ We perform pre-layer normalization, since it improves the training stability, and use the RMSNorm, which is faster.
69
+ We use the SwiGLU activation function, since it has been shown to lead to good results on downstream tasks.
70
+ We use rotary positional embeddings (RoPE) in every layer, since these have been shown to lead to good performances while allowing the extension of the context length.
71
+
72
+ For pre-training, we use 256 Nvidia H100 GPUs of the Marenostrum 5 supercomputer, training the model with a constant batch size of 3,072 sequences, which corresponds to approximately 12 million tokens, using the Adam optimizer, and BF16 precision. Here is a summary of the model hyper-parameters:
73
+
74
+ Sequence Length 4,096
75
+ Number of Layers 24
76
+ Embedding Size 2,048
77
+ FFN Hidden Size 5,632
78
+ Number of Heads 16
79
+ Number of KV Heads (GQA) 8
80
+ Activation Function SwiGLU
81
+ Position Encodings RoPE (\Theta=10,000)
82
+ Layer Norm RMSNorm
83
+ Tied Embeddings No
84
+ Embedding Parameters 0.262B
85
+ LM Head Parameters 0.262B
86
+ Non-embedding Parameters 1.133B
87
+ Total Parameters 1.657B
88
+ Run the model
89
+
90
+ from transformers import AutoModelForCausalLM, AutoTokenizer
91
+
92
+ model_id = "utter-project/EuroLLM-1.7B-Instruct"
93
+ tokenizer = AutoTokenizer.from_pretrained(model_id)
94
+ model = AutoModelForCausalLM.from_pretrained(model_id)
95
+
96
+ text = '<|im_start|>system\n<|im_end|>\n<|im_start|>user\nTranslate the following English source text to Portuguese:\nEnglish: I am a language model for european languages. \nPortuguese: <|im_end|>\n<|im_start|>assistant\n'
97
+
98
+ inputs = tokenizer(text, return_tensors="pt")
99
+ outputs = model.generate(**inputs, max_new_tokens=20)
100
+ print(tokenizer.decode(outputs[0], skip_special_tokens=True))
101
+
102
+ Results
103
+ Machine Translation
104
+
105
+ We evaluate EuroLLM-1.7B-Instruct on several machine translation benchmarks: FLORES-200, WMT-23, and WMT-24 comparing it with Gemma-2B and Gemma-7B (also instruction tuned on EuroBlocks). The results show that EuroLLM-1.7B is substantially better than Gemma-2B in Machine Translation and competitive with Gemma-7B.
106
+ Flores-200
107
+ Model AVG AVG en-xx AVG xx-en en-ar en-bg en-ca en-cs en-da en-de en-el en-es-latam en-et en-fi en-fr en-ga en-gl en-hi en-hr en-hu en-it en-ja en-ko en-lt en-lv en-mt en-nl en-no en-pl en-pt-br en-ro en-ru en-sk en-sl en-sv en-tr en-uk en-zh-cn ar-en bg-en ca-en cs-en da-en de-en el-en es-latam-en et-en fi-en fr-en ga-en gl-en hi-en hr-en hu-en it-en ja-en ko-en lt-en lv-en mt-en nl-en no-en pl-en pt-br-en ro-en ru-en sk-en sl-en sv-en tr-en uk-en zh-cn-en
108
+ EuroLLM-1.7B-Instruct 86.89 86.53 87.25 85.17 89.42 84.72 89.13 89.47 86.90 87.60 86.29 88.95 89.40 87.69 74.89 86.41 76.92 84.79 86.78 88.17 89.76 87.70 87.27 87.62 67.84 87.10 90.00 88.18 89.29 89.49 88.32 88.18 86.85 90.00 87.31 87.89 86.60 86.34 87.45 87.57 87.95 89.72 88.80 87.00 86.77 88.34 89.09 88.95 82.69 87.80 88.37 86.71 87.20 87.81 86.79 86.79 85.62 86.48 81.10 86.97 90.25 85.75 89.20 88.88 86.00 87.38 86.76 89.61 87.94
109
+ Gemma-2B-EuroBlocks 81.59 78.97 84.21 76.68 82.73 83.14 81.63 84.63 83.15 79.42 84.05 72.58 79.73 84.97 40.50 82.13 67.79 80.53 78.36 84.90 87.43 82.98 72.29 68.68 58.55 83.13 86.15 82.78 86.79 83.14 84.61 78.18 75.37 80.89 78.38 84.38 84.35 83.88 85.77 86.85 86.31 88.24 88.12 84.79 84.90 82.51 86.32 88.29 54.78 86.53 85.83 85.41 85.18 86.77 85.78 84.99 81.65 81.78 67.27 85.92 89.07 84.14 88.07 87.17 85.23 85.09 83.95 87.57 84.77
110
+ Gemma-7B-EuroBlocks 85.27 83.90 86.64 86.38 87.87 85.74 84.25 85.69 81.49 85.52 86.93 62.83 84.96 75.34 84.93 83.91 86.92 88.19 86.11 81.73 80.55 66.85 85.31 89.36 85.87 88.62 88.06 86.67 84.79 82.71 86.45 85.19 86.67 85.77 86.36 87.21 88.09 87.17 89.40 88.26 86.74 86.73 87.25 88.87 88.81 72.45 87.62 87.86 87.08 87.01 87.58 86.92 86.70 85.10 85.74 77.81 86.83 90.40 85.41 89.04 88.77 86.13 86.67 86.32 89.27 87.92
111
+ WMT-23
112
+ Model AVG AVG en-xx AVG xx-en AVG xx-xx en-de en-cs en-uk en-ru en-zh-cn de-en uk-en ru-en zh-cn-en cs-uk
113
+ EuroLLM-1.7B-Instruct 82.91 83.20 81.77 86.82 81.56 85.23 81.30 82.47 83.61 85.03 84.06 85.25 81.31 78.83
114
+ Gemma-2B-EuroBlocks 79.96 79.01 80.86 81.15 76.82 76.05 77.92 78.98 81.58 82.73 82.71 83.99 80.35 78.27
115
+ Gemma-7B-EuroBlocks 82.76 82.26 82.70 85.98 81.37 82.42 81.54 82.18 82.90 83.17 84.29 85.70 82.46 79.73
116
+ WMT-24
117
+ Model AVG AVG en-xx AVG xx-xx en-de en-es-latam en-cs en-ru en-uk en-ja en-zh-cn en-hi cs-uk ja-zh-cn
118
+ EuroLLM-1.7B-Instruct 79.32 79.32 79.34 79.42 80.67 80.55 78.65 80.12 82.96 80.60 71.59 83.48 75.20
119
+ Gemma-2B-EuroBlocks 74.72 74.41 75.97 74.93 78.81 70.54 74.90 75.84 79.48 78.06 62.70 79.87 72.07
120
+ Gemma-7B-EuroBlocks 78.67 78.34 80.00 78.88 80.47 78.55 78.55 80.12 80.55 78.90 70.71 84.33 75.66
121
+ General Benchmarks
122
+
123
+ We also compare EuroLLM-1.7B with TinyLlama-v1.1 and Gemma-2B on 3 general benchmarks: Arc Challenge and Hellaswag. For the non-english languages we use the Okapi datasets. Results show that EuroLLM-1.7B is superior to TinyLlama-v1.1 and similar to Gemma-2B on Hellaswag but worse on Arc Challenge. This can be due to the lower number of parameters of EuroLLM-1.7B (1.133B non-embedding parameters against 1.981B).
124
+ Arc Challenge
125
+ Model Average English German Spanish French Italian Portuguese Chinese Russian Dutch Arabic Swedish Hindi Hungarian Romanian Ukrainian Danish Catalan
126
+ EuroLLM-1.7B 0.3496 0.4061 0.3464 0.3684 0.3627 0.3738 0.3855 0.3521 0.3208 0.3507 0.3045 0.3605 0.2928 0.3271 0.3488 0.3516 0.3513 0.3396
127
+ TinyLlama-v1.1 0.2650 0.3712 0.2524 0.2795 0.2883 0.2652 0.2906 0.2410 0.2669 0.2404 0.2310 0.2687 0.2354 0.2449 0.2476 0.2524 0.2494 0.2796
128
+ Gemma-2B 0.3617 0.4846 0.3755 0.3940 0.4080 0.3687 0.3872 0.3726 0.3456 0.3328 0.3122 0.3519 0.2851 0.3039 0.3590 0.3601 0.3565 0.3516
129
+ Hellaswag
130
+ Model Average English German Spanish French Italian Portuguese Russian Dutch Arabic Swedish Hindi Hungarian Romanian Ukrainian Danish Catalan
131
+ EuroLLM-1.7B 0.4744 0.4760 0.6057 0.4793 0.5337 0.5298 0.5085 0.5224 0.4654 0.4949 0.4104 0.4800 0.3655 0.4097 0.4606 0.436 0.4702
132
+ TinyLlama-v1.1 0.3674 0.6248 0.3650 0.4137 0.4010 0.3780 0.3892 0.3494 0.3588 0.2880 0.3561 0.2841 0.3073 0.3267 0.3349 0.3408 0.3613
133
+ Gemma-2B 0.4666 0.7165 0.4756 0.5414 0.5180 0.4841 0.5081 0.4664 0.4655 0.3868 0.4383 0.3413 0.3710 0.4316 0.4291 0.4471 0.4448
134
+ Bias, Risks, and Limitations
135
+
136
+ EuroLLM-1.7B-Instruct has not been aligned to human preferences, so the model may generate problematic outputs (e.g., hallucinations, harmful content, or false statements).
137
+ Paper
138
+
139
+ Paper: EuroLLM: Multilingual Language Models for Europe
140
+
141
+ ---
142
  ## Use with llama.cpp
143
  Install llama.cpp through brew (works on Mac and Linux)
144