Triangle104 commited on
Commit
40937f8
1 Parent(s): 54fe7de

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +149 -0
README.md CHANGED
@@ -19,6 +19,155 @@ tags:
19
  This model was converted to GGUF format from [`tiiuae/falcon-mamba-7b-instruct`](https://huggingface.co/tiiuae/falcon-mamba-7b-instruct) using llama.cpp via the ggml.ai's [GGUF-my-repo](https://huggingface.co/spaces/ggml-org/gguf-my-repo) space.
20
  Refer to the [original model card](https://huggingface.co/tiiuae/falcon-mamba-7b-instruct) for more details on the model.
21
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
22
  ## Use with llama.cpp
23
  Install llama.cpp through brew (works on Mac and Linux)
24
 
 
19
  This model was converted to GGUF format from [`tiiuae/falcon-mamba-7b-instruct`](https://huggingface.co/tiiuae/falcon-mamba-7b-instruct) using llama.cpp via the ggml.ai's [GGUF-my-repo](https://huggingface.co/spaces/ggml-org/gguf-my-repo) space.
20
  Refer to the [original model card](https://huggingface.co/tiiuae/falcon-mamba-7b-instruct) for more details on the model.
21
 
22
+ ---
23
+ Model details:
24
+ -
25
+ Table of Contents
26
+
27
+ TL;DR
28
+ Model Details
29
+ Usage
30
+ Training Details
31
+ Evaluation
32
+
33
+ TL;DR
34
+ Model Details
35
+ Model Description
36
+
37
+ Developed by: https://www.tii.ae
38
+ Model type: Causal decoder-only
39
+ Architecture: Mamba
40
+ Language(s) (NLP): Mainly English
41
+ License: TII Falcon-Mamba License 2.0
42
+
43
+
44
+ Usage
45
+
46
+ Find below some example scripts on how to use the model in transformers (Make sure to have the latest transformers, or the one built from source):
47
+ Using the Pytorch model
48
+ Running the model on a CPU
49
+ Click to expand
50
+
51
+ Running the model on a GPU
52
+ Click to expand
53
+
54
+ Running the model on a GPU using torch.compile
55
+ Click to expand
56
+
57
+ Running the model on a GPU using different precisions
58
+ FP16
59
+ Click to expand
60
+
61
+ 4-bit
62
+ Click to expand
63
+
64
+
65
+ Training Details
66
+ Training Data
67
+
68
+ Falcon-Mamba has been trained with ~ 5,500 GT mainly coming from Refined-Web, a large volume web-only dataset filtered and deduplicated. Similar to the others Falcon suite models, Falcon-Mamba has been trained leveraging a multi-stage training strategy to increase the context-length from 2,048 to 8,192. Moreover, inspired by the concept of Curriculum Learning, we carefully selected data mixtures throughout the training stages, considering both data diversity and complexity. Note that at inference the context-length is not relevant as the Mamba architecture has no limit on long range dependency. At the last training stage, small portion of high-quality curated data was used to further enhance performance.
69
+
70
+ Overall, the data sources included RefinedWeb-English, high quality technical data, code data and math data extracted from public sources. In particular, we used samples coming from Fineweb-edu during our last training stage.
71
+
72
+ The data was tokenized with the Falcon-7B/11B tokenizer.
73
+
74
+ After pre-training, the model has been further fine-tuned on instruction data.
75
+ Training Procedure
76
+
77
+ Falcon-Mamba-7B was trained on 256 H100 80GB GPUs for the majority of the training, using a 3D parallelism strategy (TP=1, PP=1, DP=256) combined with ZeRO.
78
+ Training Hyperparameters
79
+ Hyperparameter Value Comment
80
+ Precision bfloat16
81
+ Optimizer AdamW
82
+ Max learning rate 6.4e-4 Following a WSD (warmup-stable-decay) learning rate schedule
83
+ Weight decay 1e-1
84
+ Batch size 2048
85
+
86
+ The model was trained AdamW optimizer, WSD (warmup-stable-decay) learning rate schedule, and a batch size rampup from bmin=128b_{\mathrm{min}}=128bmin​=128 to bmax=2048b_{\mathrm{max}}=2048bmax​=2048 during first 50 GT of training. In the stable phase we used maximal learning rate ηmax=6.4×10−4\eta_{\mathrm{max}}=6.4 \times 10^{-4}ηmax​=6.4×10−4, and decayed it to the minimal value ηmin=ηmax256\eta_{\mathrm{min}}=\frac{\eta_{\mathrm{max}}}{256}ηmin​=256ηmax​​ with exponential schedule over 500 GT. Also, we applied BatchScaling during the rampup — rescaling learning rate η\etaη so that the Adam noise temperature Tnoise≡ηbT_{\mathrm{noise}}\equiv\frac{\eta}{\sqrt{b}}Tnoise​≡b
87
+
88
+ ​η​ is kept constant.
89
+ Speeds, Sizes, Times
90
+
91
+ The model training took roughly two months.
92
+
93
+ Evaluation
94
+ Benchmarks
95
+
96
+ We evaluate our model on all benchmarks of the new leaderboard's version using the lm-evaluation-harness package, and then normalize the evaluation results with HuggingFace score normalization.
97
+ model name IFEval BBH MATH LvL5 GPQA MUSR MMLU-PRO Average
98
+ Pure SSM models
99
+ FalconMamba-7B 33.36 19.88 3.63 8.05 10.86 14.47 15.04
100
+ TRI-ML/mamba-7b-rw* 22.46 6.71 0.45 1.12 5.51 1.69 6.25
101
+ Hybrid SSM-attention models
102
+ recurrentgemma-9b 30.76 14.80 4.83 4.70 6.60 17.88 13.20
103
+ Zyphra/Zamba-7B-v1* 24.06 21.12 3.32 3.03 7.74 16.02 12.55
104
+ Transformer models
105
+ Falcon2-11B 32.61 21.94 2.34 2.80 7.53 15.44 13.78
106
+ Meta-Llama-3-8B 14.55 24.50 3.25 7.38 6.24 24.55 13.41
107
+ Meta-Llama-3.1-8B 12.70 25.29 4.61 6.15 8.98 24.95 13.78
108
+ Mistral-7B-v0.1 23.86 22.02 2.49 5.59 10.68 22.36 14.50
109
+ Mistral-Nemo-Base-2407 (12B) 16.83 29.37 4.98 5.82 6.52 27.46 15.08
110
+ gemma-7B 26.59 21.12 6.42 4.92 10.98 21.64 15.28
111
+
112
+ Also, we evaluate our model on the benchmarks of the first leaderboard using lighteval.
113
+ model name ARC HellaSwag MMLU Winogrande TruthfulQA GSM8K Average
114
+ Pure SSM models
115
+ FalconMamba-7B* 62.03 80.82 62.11 73.64 53.42 52.54 64.09
116
+ TRI-ML/mamba-7b-rw* 51.25 80.85 33.41 71.11 32.08 4.70 45.52
117
+ Hybrid SSM-attention models
118
+ recurrentgemma-9b** 52.00 80.40 60.50 73.60 38.60 42.60 57.95
119
+ Zyphra/Zamba-7B-v1* 56.14 82.23 58.11 79.87 52.88 30.78 60.00
120
+ Transformer models
121
+ Falcon2-11B 59.73 82.91 58.37 78.30 52.56 53.83 64.28
122
+ Meta-Llama-3-8B 60.24 82.23 66.70 78.45 42.93 45.19 62.62
123
+ Meta-Llama-3.1-8B 58.53 82.13 66.43 74.35 44.29 47.92 62.28
124
+ Mistral-7B-v0.1 59.98 83.31 64.16 78.37 42.15 37.83 60.97
125
+ gemma-7B 61.09 82.20 64.56 79.01 44.79 50.87 63.75
126
+
127
+ Mostly, we took evaluation results from both leaderboards. For the models marked by star we evaluated the tasks internally, while for the models marked by two stars the results were taken from paper or model card.
128
+ Throughput
129
+
130
+ This model can achieve comparable throughput and performance compared to other transformer based models that use optimized kernels such as Flash Attention 2. Make sure to install the optimized Mamba kernels with the following commands:
131
+
132
+ pip install "causal-conv1d>=1.4.0" mamba-ssm
133
+
134
+ Refer to our FalconMamba blogpost for more details about performance evaluation.
135
+
136
+ Technical Specifications
137
+ Model Architecture and Objective
138
+
139
+ Falcon-Mamba-7B is a causal decoder-only model trained on a causal language modeling task (i.e., predict the next token).
140
+
141
+ The model is based on the Mamba architecture (Gu et al., 2023).
142
+ Hyperparameter Value Comment
143
+ Layers 64 Number of layers
144
+ d_model 4096 Hidden dimension
145
+ d_state 16 The SSM state dimension
146
+ Vocabulary 65024 Vocabulary Size
147
+ Sequence length 8192 During the last training stages
148
+ Compute Infrastructure
149
+ Hardware
150
+
151
+ Falcon-Mamba-7B was trained on AWS SageMaker, using on average 256 H100 80GB GPUs in 32 p5 instances.
152
+ Software
153
+
154
+ Falcon-Mamba-7B was trained on an internal distributed training codebase, Gigatron. It uses a 3D parallelism approach combined with ZeRO, high-performance Triton kernels.
155
+
156
+ Citation
157
+
158
+ You can use the following bibtex citation:
159
+
160
+ @misc{zuo2024falconmambacompetitiveattentionfree,
161
+ title={Falcon Mamba: The First Competitive Attention-free 7B Language Model},
162
+ author={Jingwei Zuo and Maksim Velikanov and Dhia Eddine Rhaiem and Ilyas Chahed and Younes Belkada and Guillaume Kunsch and Hakim Hacid},
163
+ year={2024},
164
+ eprint={2410.05355},
165
+ archivePrefix={arXiv},
166
+ primaryClass={cs.CL},
167
+ url={https://arxiv.org/abs/2410.05355},
168
+ }
169
+
170
+ ---
171
  ## Use with llama.cpp
172
  Install llama.cpp through brew (works on Mac and Linux)
173