README.md
8.2 KB · 173 lines · markdown Raw
1 ---
2 library_name: transformers
3 license: apache-2.0
4 language:
5 - en
6 tags:
7 - fill-mask
8 - masked-lm
9 - long-context
10 - modernbert
11 pipeline_tag: fill-mask
12 inference: false
13 ---
14
15 # ModernBERT
16
17 ## Table of Contents
18 1. [Model Summary](#model-summary)
19 2. [Usage](#Usage)
20 3. [Evaluation](#Evaluation)
21 4. [Limitations](#limitations)
22 5. [Training](#training)
23 6. [License](#license)
24 7. [Citation](#citation)
25
26 ## Model Summary
27
28 ModernBERT is a modernized bidirectional encoder-only Transformer model (BERT-style) pre-trained on 2 trillion tokens of English and code data with a native context length of up to 8,192 tokens. ModernBERT leverages recent architectural improvements such as:
29
30 - **Rotary Positional Embeddings (RoPE)** for long-context support.
31 - **Local-Global Alternating Attention** for efficiency on long inputs.
32 - **Unpadding and Flash Attention** for efficient inference.
33
34 ModernBERT’s native long context length makes it ideal for tasks that require processing long documents, such as retrieval, classification, and semantic search within large corpora. The model was trained on a large corpus of text and code, making it suitable for a wide range of downstream tasks, including code retrieval and hybrid (text + code) semantic search.
35
36 It is available in the following sizes:
37
38 - [ModernBERT-base](https://huggingface.co/answerdotai/ModernBERT-base) - 22 layers, 149 million parameters
39 - [ModernBERT-large](https://huggingface.co/answerdotai/ModernBERT-large) - 28 layers, 395 million parameters
40
41 For more information about ModernBERT, we recommend our [release blog post](https://huggingface.co/blog/modernbert) for a high-level overview, and our [arXiv pre-print](https://arxiv.org/abs/2412.13663) for in-depth information.
42
43 *ModernBERT is a collaboration between [Answer.AI](https://answer.ai), [LightOn](https://lighton.ai), and friends.*
44
45 ## Usage
46
47 You can use these models directly with the `transformers` library starting from v4.48.0:
48
49 ```sh
50 pip install -U transformers>=4.48.0
51 ```
52
53 Since ModernBERT is a Masked Language Model (MLM), you can use the `fill-mask` pipeline or load it via `AutoModelForMaskedLM`. To use ModernBERT for downstream tasks like classification, retrieval, or QA, fine-tune it following standard BERT fine-tuning recipes.
54
55 **⚠️ If your GPU supports it, we recommend using ModernBERT with Flash Attention 2 to reach the highest efficiency. To do so, install Flash Attention as follows, then use the model as normal:**
56
57 ```bash
58 pip install flash-attn
59 ```
60
61 Using `AutoModelForMaskedLM`:
62
63 ```python
64 from transformers import AutoTokenizer, AutoModelForMaskedLM
65
66 model_id = "answerdotai/ModernBERT-base"
67 tokenizer = AutoTokenizer.from_pretrained(model_id)
68 model = AutoModelForMaskedLM.from_pretrained(model_id)
69
70 text = "The capital of France is [MASK]."
71 inputs = tokenizer(text, return_tensors="pt")
72 outputs = model(**inputs)
73
74 # To get predictions for the mask:
75 masked_index = inputs["input_ids"][0].tolist().index(tokenizer.mask_token_id)
76 predicted_token_id = outputs.logits[0, masked_index].argmax(axis=-1)
77 predicted_token = tokenizer.decode(predicted_token_id)
78 print("Predicted token:", predicted_token)
79 # Predicted token: Paris
80 ```
81
82 Using a pipeline:
83
84 ```python
85 import torch
86 from transformers import pipeline
87 from pprint import pprint
88
89 pipe = pipeline(
90 "fill-mask",
91 model="answerdotai/ModernBERT-base",
92 torch_dtype=torch.bfloat16,
93 )
94
95 input_text = "He walked to the [MASK]."
96 results = pipe(input_text)
97 pprint(results)
98 ```
99
100 **Note:** ModernBERT does not use token type IDs, unlike some earlier BERT models. Most downstream usage is identical to standard BERT models on the Hugging Face Hub, except you can omit the `token_type_ids` parameter.
101
102 ## Evaluation
103
104 We evaluate ModernBERT across a range of tasks, including natural language understanding (GLUE), general retrieval (BEIR), long-context retrieval (MLDR), and code retrieval (CodeSearchNet and StackQA).
105
106 **Key highlights:**
107 - On GLUE, ModernBERT-base surpasses other similarly-sized encoder models, and ModernBERT-large is second only to Deberta-v3-large.
108 - For general retrieval tasks, ModernBERT performs well on BEIR in both single-vector (DPR-style) and multi-vector (ColBERT-style) settings.
109 - Thanks to the inclusion of code data in its training mixture, ModernBERT as a backbone also achieves new state-of-the-art code retrieval results on CodeSearchNet and StackQA.
110
111 ### Base Models
112
113 | Model | IR (DPR) | IR (DPR) | IR (DPR) | IR (ColBERT) | IR (ColBERT) | NLU | Code | Code |
114 |-------------|--------------|--------------|--------------|---------------|---------------|------|------|------|
115 | | BEIR | MLDR_OOD | MLDR_ID | BEIR | MLDR_OOD | GLUE | CSN | SQA |
116 | BERT | 38.9 | 23.9 | 32.2 | 49.0 | 28.1 | 84.7 | 41.2 | 59.5 |
117 | RoBERTa | 37.7 | 22.9 | 32.8 | 48.7 | 28.2 | 86.4 | 44.3 | 59.6 |
118 | DeBERTaV3 | 20.2 | 5.4 | 13.4 | 47.1 | 21.9 | 88.1 | 17.5 | 18.6 |
119 | NomicBERT | 41.0 | 26.7 | 30.3 | 49.9 | 61.3 | 84.0 | 41.6 | 61.4 |
120 | GTE-en-MLM | 41.4 | **34.3** |**44.4** | 48.2 | 69.3 | 85.6 | 44.9 | 71.4 |
121 | ModernBERT | **41.6** | 27.4 | 44.0 | **51.3** | **80.2** | **88.4** | **56.4** |**73.6**|
122
123 ---
124
125 ### Large Models
126
127 | Model | IR (DPR) | IR (DPR) | IR (DPR) | IR (ColBERT) | IR (ColBERT) | NLU | Code | Code |
128 |-------------|--------------|--------------|--------------|---------------|---------------|------|------|------|
129 | | BEIR | MLDR_OOD | MLDR_ID | BEIR | MLDR_OOD | GLUE | CSN | SQA |
130 | BERT | 38.9 | 23.3 | 31.7 | 49.5 | 28.5 | 85.2 | 41.6 | 60.8 |
131 | RoBERTa | 41.4 | 22.6 | 36.1 | 49.8 | 28.8 | 88.9 | 47.3 | 68.1 |
132 | DeBERTaV3 | 25.6 | 7.1 | 19.2 | 46.7 | 23.0 | **91.4**| 21.2 | 19.7 |
133 | GTE-en-MLM | 42.5 | **36.4** | **48.9** | 50.7 | 71.3 | 87.6 | 40.5 | 66.9 |
134 | ModernBERT | **44.0** | 34.3 | 48.6 | **52.4** | **80.4** | 90.4 |**59.5** |**83.9**|
135
136 *Table 1: Results for all models across an overview of all tasks. CSN refers to CodeSearchNet and SQA to StackQA. MLDRID refers to in-domain (fine-tuned on the training set) evaluation, and MLDR_OOD to out-of-domain.*
137
138 ModernBERT’s strong results, coupled with its efficient runtime on long-context inputs, demonstrate that encoder-only models can be significantly improved through modern architectural choices and extensive pretraining on diversified data sources.
139
140
141 ## Limitations
142
143 ModernBERT’s training data is primarily English and code, so performance may be lower for other languages. While it can handle long sequences efficiently, using the full 8,192 tokens window may be slower than short-context inference. Like any large language model, ModernBERT may produce representations that reflect biases present in its training data. Verify critical or sensitive outputs before relying on them.
144
145 ## Training
146
147 - Architecture: Encoder-only, Pre-Norm Transformer with GeGLU activations.
148 - Sequence Length: Pre-trained up to 1,024 tokens, then extended to 8,192 tokens.
149 - Data: 2 trillion tokens of English text and code.
150 - Optimizer: StableAdamW with trapezoidal LR scheduling and 1-sqrt decay.
151 - Hardware: Trained on 8x H100 GPUs.
152
153 See the paper for more details.
154
155 ## License
156
157 We release the ModernBERT model architectures, model weights, training codebase under the Apache 2.0 license.
158
159 ## Citation
160
161 If you use ModernBERT in your work, please cite:
162
163 ```
164 @misc{modernbert,
165 title={Smarter, Better, Faster, Longer: A Modern Bidirectional Encoder for Fast, Memory Efficient, and Long Context Finetuning and Inference},
166 author={Benjamin Warner and Antoine Chaffin and Benjamin Clavié and Orion Weller and Oskar Hallström and Said Taghadouini and Alexis Gallagher and Raja Biswas and Faisal Ladhak and Tom Aarsen and Nathan Cooper and Griffin Adams and Jeremy Howard and Iacopo Poli},
167 year={2024},
168 eprint={2412.13663},
169 archivePrefix={arXiv},
170 primaryClass={cs.CL},
171 url={https://arxiv.org/abs/2412.13663},
172 }
173 ```