GGUF
English
Inference Endpoints
conversational
ybelkada commited on
Commit
70679d5
1 Parent(s): 3d48281

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +182 -0
README.md CHANGED
@@ -4,3 +4,185 @@ license_name: falcon-mamba-license
4
  license_link: https://falconllm.tii.ae/falcon-mamba-7b-terms-and-conditions.html
5
  base_model: tiiuae/falcon-mamba-7b-instruct
6
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4
  license_link: https://falconllm.tii.ae/falcon-mamba-7b-terms-and-conditions.html
5
  base_model: tiiuae/falcon-mamba-7b-instruct
6
  ---
7
+ ---
8
+ language:
9
+ - en
10
+ datasets:
11
+ - tiiuae/falcon-refinedweb
12
+ - HuggingFaceFW/fineweb-edu
13
+ license: other
14
+ license_name: falcon-mamba-7b-license
15
+ license_link: https://falconllm.tii.ae/falcon-mamba-7b-terms-and-conditions.html
16
+ ---
17
+
18
+ <img src="https://huggingface.co/datasets/tiiuae/documentation-images/resolve/main/falcon_mamba/thumbnail.png" alt="drawing" width="800"/>
19
+
20
+ **GGUF quantization of [`falcon-mamba-7b-instruct`](https://huggingface.co/tiiuae/falcon-mamba-7b-instruct)**
21
+
22
+ # Table of Contents
23
+
24
+ 0. [TL;DR](#TL;DR)
25
+ 1. [Model Details](#model-details)
26
+ 2. [Usage](#usage)
27
+ 3. [Training Details](#training-details)
28
+ 4. [Evaluation](#evaluation)
29
+
30
+
31
+ # TL;DR
32
+
33
+ # Model Details
34
+
35
+ ## Model Description
36
+
37
+ - **Developed by:** [https://www.tii.ae](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
+ <br>
44
+
45
+ # Usage
46
+
47
+ Refer to the documentation of [`llama.cpp`](https://github.com/ggerganov/llama.cpp) to understand how to run this model locally on your machine.
48
+
49
+ # Training Details
50
+
51
+ ## Training Data
52
+
53
+ Falcon-Mamba has been trained with ~ 5,500 GT mainly coming from [Refined-Web](https://huggingface.co/datasets/tiiuae/falcon-refinedweb), a large volume web-only dataset filtered and deduplicated.
54
+ Similar to the others [Falcon](https://huggingface.co/tiiuae/falcon-11B) suite models, Falcon-Mamba has been trained leveraging a multi-stage training strategy to increase the context-length from 2,048 to 8,192.
55
+ Moreover, inspired by the concept of Curriculum Learning, we carefully selected data mixtures throughout the training stages, considering both data diversity and complexity.
56
+ Note that at inference the context-length is not relevant as the Mamba architecture has no limit on long range dependency.
57
+ At the last training stage, small portion of high-quality curated data was used to further enhance performance.
58
+
59
+ Overall, the data sources included RefinedWeb-English, high quality technical data, code data and math data extracted from public sources.
60
+ In particular, we used samples coming from [Fineweb-edu](https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu) during our last training stage.
61
+
62
+ The data was tokenized with the Falcon-[7B](https://huggingface.co/tiiuae/falcon-7B)/[11B](https://huggingface.co/tiiuae/falcon-11B) tokenizer.
63
+
64
+ ## Training Procedure
65
+ 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.
66
+
67
+ ### Training Hyperparameters
68
+
69
+ | **Hyperparameter** | **Value** | **Comment** |
70
+ |--------------------|------------|-------------------------------------------|
71
+ | Precision | `bfloat16` | |
72
+ | Optimizer | AdamW | |
73
+ | Max learning rate | 6.4e-4 | Following a WSD (warmup-stable-decay) learning rate schedule |
74
+ | Weight decay | 1e-1 | |
75
+ | Batch size | 2048 | |
76
+
77
+
78
+ The model was trained AdamW optimizer, WSD (warmup-stable-decay) learning rate schedule, and a batch size rampup from \\(b_{\mathrm{min}}=128\\) to \\(b_{\mathrm{max}}=2048\\) during first 50 GT of training.
79
+ In the stable phase we used maximal learning rate \\(\eta_{\mathrm{max}}=6.4 \times 10^{-4}\\), and decayed it to the minimal value \\(\eta_{\mathrm{min}}=\frac{\eta_{\mathrm{max}}}{256}\\) with exponential schedule over 500 GT.
80
+ Also, we applied *BatchScaling* during the rampup — rescaling learning rate \\(\eta\\) so that the Adam noise temperature \\(T_{\mathrm{noise}}\equiv\frac{\eta}{\sqrt{b}}\\) is kept constant.
81
+
82
+ ### Speeds, Sizes, Times
83
+
84
+ The model training took roughly two months.
85
+
86
+ <br>
87
+
88
+ # Evaluation
89
+
90
+ ## Benchmarks
91
+
92
+ 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.
93
+
94
+
95
+ | `model name` |`IFEval`| `BBH` |`MATH LvL5`| `GPQA`| `MUSR`|`MMLU-PRO`|`Average`|
96
+ |:--------------------------|:------:|:-----:|:---------:|:-----:|:-----:|:--------:|:-------:|
97
+ | ***Pure SSM models*** | | | | | | | |
98
+ | `FalconMamba-7B` | 33.36 | 19.88 | 3.63 |8.05 |10.86 | 14.47 |**15.04**|
99
+ | `TRI-ML/mamba-7b-rw`<sup>*</sup>| 22.46 | 6.71 | 0.45 | 1.12 | 5.51 | 1.69 | 6.25 |
100
+ |***Hybrid SSM-attention models*** | | | | | | |
101
+ |`recurrentgemma-9b` | 30.76 | 14.80 | 4.83 | 4.70 | 6.60 | 17.88 | 13.20 |
102
+ | `Zyphra/Zamba-7B-v1`<sup>*</sup> | 24.06 | 21.12 | 3.32 | 3.03 | 7.74 | 16.02 | 12.55 |
103
+ |***Transformer models*** | | | | | | | |
104
+ | `Falcon2-11B` | 32.61 | 21.94 | 2.34 | 2.80 | 7.53 | 15.44 | 13.78 |
105
+ | `Meta-Llama-3-8B` | 14.55 | 24.50 | 3.25 | 7.38 | 6.24 | 24.55 | 13.41 |
106
+ | `Meta-Llama-3.1-8B` | 12.70 | 25.29 | 4.61 | 6.15 | 8.98 | 24.95 | 13.78 |
107
+ | `Mistral-7B-v0.1` | 23.86 | 22.02 | 2.49 | 5.59 | 10.68 | 22.36 | 14.50 |
108
+ | `Mistral-Nemo-Base-2407 (12B)` | 16.83 | 29.37 | 4.98 | 5.82 | 6.52 | 27.46 | 15.08 |
109
+ | `gemma-7B` | 26.59 | 21.12 | 6.42 | 4.92 | 10.98 | 21.64 |**15.28**|
110
+
111
+
112
+ Also, we evaluate our model on the benchmarks of the first leaderboard using `lighteval`.
113
+
114
+
115
+ | `model name` |`ARC`|`HellaSwag` |`MMLU` |`Winogrande`|`TruthfulQA`|`GSM8K`|`Average` |
116
+ |:-----------------------------|:------:|:---------:|:-----:|:----------:|:----------:|:-----:|:----------------:|
117
+ | ***Pure SSM models*** | | | | | | | |
118
+ | `FalconMamba-7B`<sup>*</sup> | 62.03 | 80.82 | 62.11 | 73.64 | 53.42 | 52.54 | **64.09** |
119
+ | `TRI-ML/mamba-7b-rw`<sup>*</sup> | 51.25 | 80.85 | 33.41 | 71.11 | 32.08 | 4.70 | 45.52 |
120
+ |***Hybrid SSM-attention models***| | | | | | | |
121
+ | `recurrentgemma-9b`<sup>**</sup> |52.00 | 80.40 | 60.50 | 73.60 | 38.60 | 42.60 | 57.95 |
122
+ | `Zyphra/Zamba-7B-v1`<sup>*</sup> | 56.14 | 82.23 | 58.11 | 79.87 | 52.88 | 30.78 | 60.00 |
123
+ |***Transformer models*** | | | | | | | |
124
+ | `Falcon2-11B` | 59.73 | 82.91 | 58.37 | 78.30 | 52.56 | 53.83 | **64.28** |
125
+ | `Meta-Llama-3-8B` | 60.24 | 82.23 | 66.70 | 78.45 | 42.93 | 45.19 | 62.62 |
126
+ | `Meta-Llama-3.1-8B` | 58.53 | 82.13 | 66.43 | 74.35 | 44.29 | 47.92 | 62.28 |
127
+ | `Mistral-7B-v0.1` | 59.98 | 83.31 | 64.16 | 78.37 | 42.15 | 37.83 | 60.97 |
128
+ | `gemma-7B` | 61.09 | 82.20 | 64.56 | 79.01 | 44.79 | 50.87 | 63.75 |
129
+
130
+ 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.
131
+
132
+ ## Throughput
133
+
134
+ 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:
135
+
136
+ ```bash
137
+ pip install "causal-conv1d>=1.4.0" mamba-ssm
138
+ ```
139
+
140
+ Refer to our [FalconMamba blogpost](https://huggingface.co/blog/falconmamba) for more details about performance evaluation.
141
+
142
+
143
+ <br>
144
+
145
+ # Technical Specifications
146
+
147
+ ## Model Architecture and Objective
148
+
149
+ Falcon-Mamba-7B is a causal decoder-only model trained on a causal language modeling task (i.e., predict the next token).
150
+
151
+ The model is based on the Mamba architecture ([Gu et al., 2023](https://arxiv.org/abs/2312.00752)).
152
+
153
+ | **Hyperparameter** | **Value** | **Comment** |
154
+ |--------------------|-----------|----------------------------------------|
155
+ | Layers | 64 | Number of layers |
156
+ | `d_model` | 4096 | Hidden dimension |
157
+ | `d_state` | 16 | The SSM state dimension |
158
+ | Vocabulary | 65024 | Vocabulary Size |
159
+ | Sequence length | 8192 | During the last training stages |
160
+
161
+ ## Compute Infrastructure
162
+
163
+ ### Hardware
164
+
165
+ Falcon-Mamba-7B was trained on AWS SageMaker, using on average 256 H100 80GB GPUs in 32 p5 instances.
166
+
167
+ ### Software
168
+
169
+ 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.
170
+
171
+ <br>
172
+
173
+ # Citation
174
+
175
+ *Paper coming soon* 😊.
176
+ # [Open LLM Leaderboard Evaluation Results](https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard)
177
+ Detailed results can be found [here](https://huggingface.co/datasets/open-llm-leaderboard/tiiuae__falcon-mamba-7b-details)
178
+
179
+ | Metric |Value|
180
+ |-------------------|----:|
181
+ |Avg. |15.04|
182
+ |IFEval (0-Shot) |33.36|
183
+ |BBH (3-Shot) |19.88|
184
+ |MATH Lvl 5 (4-Shot)| 3.63|
185
+ |GPQA (0-shot) | 8.05|
186
+ |MuSR (0-shot) |10.86|
187
+ |MMLU-PRO (5-shot) |14.47|
188
+