Update README.md

README.md

@@ -1,199 +1,295 @@
 ---
-library_name: transformers
-tags: []
+language:
+- en
+datasets:
+- tiiuae/falcon-refinedweb
+- HuggingFaceFW/fineweb-edu
+license: other
+license_name: falcon-mamba-7b-license
+license_link: https://falconllm.tii.ae/falcon-mamba-7b-terms-and-conditions.html
 ---
 
-# Model Card for Model ID
+<img src="https://huggingface.co/datasets/tiiuae/documentation-images/resolve/main/falcon_mamba/thumbnail.png" alt="drawing" width="800"/>
 
-<!-- Provide a quick summary of what the model is/does. -->
+#  Table of Contents
 
+0. [TL;DR](#TL;DR)
+1. [Model Details](#model-details)
+2. [Usage](#usage)
+3. [Training Details](#training-details)
+4. [Evaluation](#evaluation)
 
+Falcon Mamba 7B - pre-decay checkpoint for continuous pretraining.
 
-## Model Details
+# TL;DR
 
-### Model Description
+# Model Details
 
-<!-- Provide a longer summary of what this model is. -->
+## Model Description
 
-This is the model card of a 🤗 transformers model that has been pushed on the Hub. This model card has been automatically generated.
+- **Developed by:** [https://www.tii.ae](https://www.tii.ae)
+- **Model type:** Causal decoder-only
+- **Architecture:** Mamba
+- **Language(s) (NLP):** Mainly English
+- **License:** TII Falcon-Mamba License 2.0
 
-- **Developed by:** [More Information Needed]
-- **Funded by [optional]:** [More Information Needed]
-- **Shared by [optional]:** [More Information Needed]
-- **Model type:** [More Information Needed]
-- **Language(s) (NLP):** [More Information Needed]
-- **License:** [More Information Needed]
-- **Finetuned from model [optional]:** [More Information Needed]
+<br>
 
-### Model Sources [optional]
+# Usage
 
-<!-- Provide the basic links for the model. -->
+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):
 
-- **Repository:** [More Information Needed]
-- **Paper [optional]:** [More Information Needed]
-- **Demo [optional]:** [More Information Needed]
+## Using the Pytorch model
 
-## Uses
+### Running the model on a CPU
 
-<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+<details>
+<summary> Click to expand </summary>
 
-### Direct Use
+```python
+from transformers import AutoTokenizer, AutoModelForCausalLM
 
-<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+tokenizer = AutoTokenizer.from_pretrained("tiiuae/falcon-mamba-7b-pre-decay")
+model = AutoModelForCausalLM.from_pretrained("tiiuae/falcon-mamba-7b-pre-decay")
 
-[More Information Needed]
+input_text = "Question: How many hours in one day? Answer: "
+input_ids = tokenizer(input_text, return_tensors="pt").input_ids
 
-### Downstream Use [optional]
+outputs = model.generate(input_ids)
+print(tokenizer.decode(outputs[0]))
+```
 
-<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+</details>
 
-[More Information Needed]
+### Running the model on a GPU
 
-### Out-of-Scope Use
+<details>
+<summary> Click to expand </summary>
 
-<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+```python
+# pip install accelerate
+from transformers import AutoTokenizer, AutoModelForCausalLM
 
-[More Information Needed]
+tokenizer = AutoTokenizer.from_pretrained("tiiuae/falcon-mamba-7b-pre-decay")
+model = AutoModelForCausalLM.from_pretrained("tiiuae/falcon-mamba-7b-pre-decay", device_map="auto")
 
-## Bias, Risks, and Limitations
+input_text = "Question: How many hours in one day? Answer: "
+input_ids = tokenizer(input_text, return_tensors="pt").input_ids.to("cuda")
 
-<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+outputs = model.generate(input_ids)
+print(tokenizer.decode(outputs[0]))
+```
 
-[More Information Needed]
+</details>
 
-### Recommendations
+### Running the model on a GPU using `torch.compile`
 
-<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+<details>
+<summary> Click to expand </summary>
 
-Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+```python
+import torch
+from transformers import AutoTokenizer, AutoModelForCausalLM
 
-## How to Get Started with the Model
+tokenizer = AutoTokenizer.from_pretrained("tiiuae/falcon-mamba-7b-pre-decay")
+model = AutoModelForCausalLM.from_pretrained("tiiuae/falcon-mamba-7b-pre-decay", torch_dtype=torch.bfloat16).to(0)
 
-Use the code below to get started with the model.
+model = torch.compile(model)
 
-[More Information Needed]
+input_text = "Question: How many hours in one day? Answer: "
+input_ids = tokenizer(input_text, return_tensors="pt").input_ids.to("cuda")
 
-## Training Details
+outputs = model.generate(input_ids)
+print(tokenizer.decode(outputs[0]))
+```
 
-### Training Data
+</details>
 
-<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
 
-[More Information Needed]
+### Running the model on a GPU using different precisions
 
-### Training Procedure
+#### FP16
 
-<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+<details>
+<summary> Click to expand </summary>
 
-#### Preprocessing [optional]
+```python
+# pip install accelerate
+import torch
+from transformers import AutoTokenizer, AutoModelForCausalLM
 
-[More Information Needed]
+tokenizer = AutoTokenizer.from_pretrained("tiiuae/falcon-mamba-7b-pre-decay")
+model = AutoModelForCausalLM.from_pretrained("tiiuae/falcon-mamba-7b-pre-decay", device_map="auto", torch_dtype=torch.float16)
 
+input_text = "Question: How many hours in one day? Answer: "
+input_ids = tokenizer(input_text, return_tensors="pt").input_ids.to("cuda")
 
-#### Training Hyperparameters
+outputs = model.generate(input_ids)
+print(tokenizer.decode(outputs[0]))
+```
 
-- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+</details>
 
-#### Speeds, Sizes, Times [optional]
+#### 4-bit
 
-<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+<details>
+<summary> Click to expand </summary>
 
-[More Information Needed]
+```python
+# pip install bitsandbytes accelerate
+from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
 
-## Evaluation
+tokenizer = AutoTokenizer.from_pretrained("tiiuae/falcon-mamba-7b-pre-decay")
+model = AutoModelForCausalLM.from_pretrained("tiiuae/falcon-mamba-7b-pre-decay", device_map="auto", quantization_config=BitsAndBytesConfig(load_in_4bit=True))
 
-<!-- This section describes the evaluation protocols and provides the results. -->
+input_text = "Question: How many hours in one day? Answer: "
+input_ids = tokenizer(input_text, return_tensors="pt").input_ids.to("cuda")
 
-### Testing Data, Factors & Metrics
+outputs = model.generate(input_ids)
+print(tokenizer.decode(outputs[0]))
+```
 
-#### Testing Data
+</details>
 
-<!-- This should link to a Dataset Card if possible. -->
+<br>
 
-[More Information Needed]
+# Training Details
 
-#### Factors
+## Training Data
 
-<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+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.
+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. 
+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.
 
-[More Information Needed]
+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](https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu) during our last training stage.
 
-#### Metrics
+The data was tokenized with the Falcon-[7B](https://huggingface.co/tiiuae/falcon-7B)/[11B](https://huggingface.co/tiiuae/falcon-11B) tokenizer.
 
-<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+## Training Procedure
+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.
 
-[More Information Needed]
+### Training Hyperparameters
 
-### Results
+| **Hyperparameter** | **Value**  | **Comment**                               |
+|--------------------|------------|-------------------------------------------|
+| Precision          | `bfloat16` |                                           |
+| Optimizer          | AdamW      |                                           |
+| Max learning rate  | 6.4e-4     | Following a WSD (warmup-stable-decay) learning rate schedule |
+| Weight decay       | 1e-1       |                                           |
+| Batch size         | 2048       |                                           |
 
-[More Information Needed]
 
-#### Summary
+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. 
+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. 
+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.  
 
+### Speeds, Sizes, Times
 
+The model training took roughly two months. 
 
-## Model Examination [optional]
+<br>
 
-<!-- Relevant interpretability work for the model goes here -->
+# Evaluation
 
-[More Information Needed]
+## Benchmarks
 
-## Environmental Impact
+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.
 
-<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
 
-Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+| `model name`              |`IFEval`| `BBH` |`MATH LvL5`| `GPQA`| `MUSR`|`MMLU-PRO`|`Average`| 
+|:--------------------------|:------:|:-----:|:---------:|:-----:|:-----:|:--------:|:-------:|
+| ***Pure SSM models***     |        |       |           |       |       |          |         |
+| `FalconMamba-7B`          |  33.36 | 19.88 |    3.63   |8.05   |10.86  | 14.47    |**15.04**|
+| `TRI-ML/mamba-7b-rw`<sup>*</sup>| 22.46  | 6.71  | 0.45      | 1.12  | 5.51  | 1.69     | 6.25    |
+|***Hybrid SSM-attention models***   |       |           |       |       |          |         |
+|`recurrentgemma-9b`        | 30.76  | 14.80 | 4.83      | 4.70  | 6.60  | 17.88    |  13.20  |
+| `Zyphra/Zamba-7B-v1`<sup>*</sup>      | 24.06  | 21.12 | 3.32      | 3.03  | 7.74  | 16.02    | 12.55   |
+|***Transformer models***   |        |       |           |       |       |          |         |
+| `Falcon2-11B`             | 32.61  | 21.94 |    2.34   | 2.80  | 7.53  | 15.44    |  13.78  |
+| `Meta-Llama-3-8B`         | 14.55  | 24.50 |    3.25   | 7.38  | 6.24  | 24.55    |  13.41  |
+| `Meta-Llama-3.1-8B`       | 12.70  | 25.29 |    4.61   | 6.15  | 8.98  | 24.95    |  13.78  |
+| `Mistral-7B-v0.1`         | 23.86  | 22.02 |    2.49   | 5.59  | 10.68 | 22.36    |  14.50  |
+| `Mistral-Nemo-Base-2407 (12B)`       | 16.83  | 29.37 |    4.98   | 5.82  | 6.52  | 27.46    |  15.08  |
+| `gemma-7B`                | 26.59  | 21.12 |    6.42   | 4.92  | 10.98 | 21.64    |**15.28**|
+|***RWKV models***          |        |       |           |       |       |          |         |
+| `RWKV-v6-Finch-7B`<sup>*</sup>          | 27.65  | 9.04 |    1.11   | 2.81  | 2.25  | 5.85    |  8.12  |
+| `RWKV-v6-Finch-14B`<sup>*</sup>         | 29.81  | 12.89 |    1.13   | 5.01  | 3.16  | 11.3    |  10.55  |
 
-- **Hardware Type:** [More Information Needed]
-- **Hours used:** [More Information Needed]
-- **Cloud Provider:** [More Information Needed]
-- **Compute Region:** [More Information Needed]
-- **Carbon Emitted:** [More Information Needed]
+Also, we evaluate our model on the benchmarks of the first leaderboard using `lighteval`.
 
-## Technical Specifications [optional]
 
-### Model Architecture and Objective
+| `model name`                 |`ARC`|`HellaSwag`   |`MMLU` |`Winogrande`|`TruthfulQA`|`GSM8K`|`Average`         | 
+|:-----------------------------|:------:|:---------:|:-----:|:----------:|:----------:|:-----:|:----------------:|
+| ***Pure SSM models***        |        |           |       |            |            |       |                  |
+| `FalconMamba-7B`<sup>*</sup>          | 62.03 |   80.82   | 62.11 |   73.64    |  53.42  | 52.54 |  **64.09**       |
+| `TRI-ML/mamba-7b-rw`<sup>*</sup>         | 51.25  | 80.85     | 33.41 | 71.11      | 32.08      | 4.70  | 45.52            |
+|***Hybrid SSM-attention models***|     |           |       |            |            |       |                  |
+| `recurrentgemma-9b`<sup>**</sup>          |52.00   |   80.40   | 60.50 |   73.60    |   38.60    | 42.60 |  57.95           |
+| `Zyphra/Zamba-7B-v1`<sup>*</sup>         | 56.14  | 82.23     | 58.11 | 79.87      | 52.88      | 30.78 |  60.00           |
+|***Transformer models***      |        |           |       |            |            |       |                  |
+| `Falcon2-11B`                | 59.73  | 82.91     | 58.37 | 78.30      | 52.56      | 53.83 | **64.28**        |
+| `Meta-Llama-3-8B`            | 60.24  | 82.23     | 66.70 | 78.45      | 42.93      | 45.19 | 62.62            |
+| `Meta-Llama-3.1-8B`            | 58.53  | 82.13     | 66.43 | 74.35      | 44.29      | 47.92 | 62.28            |
+| `Mistral-7B-v0.1`            | 59.98  | 83.31     | 64.16 | 78.37      | 42.15      | 37.83 | 60.97            |
+| `Mistral-Nemo-Base-2407 (12B)`<sup>*</sup>       | 57.94  | 82.82 |    64.43   | 73.72  | 49.14  | 55.27    |  63.89  |
+| `gemma-7B`                   | 61.09  |   82.20   | 64.56 |   79.01    |   44.79    | 50.87 |  63.75           |
+|***RWKV models***             |        |       |           |       |       |          |         |
+| `RWKV-v6-Finch-7B`<sup>*</sup>          | 43.86  | 75.19 |    41.69   | 68.27  | 42.19  | 19.64    |  48.47  |
+| `RWKV-v6-Finch-14B`<sup>*</sup>         | 47.44  | 78.86 |    52.33   | 71.27  | 45.45  | 38.06    |  55.57  |
 
-[More Information Needed]
+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.
 
-### Compute Infrastructure
+## Throughput
 
-[More Information Needed]
+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:
 
-#### Hardware
+```bash
+pip install "causal-conv1d>=1.4.0" mamba-ssm
+```
 
-[More Information Needed]
+Refer to our [FalconMamba blogpost](https://huggingface.co/blog/falconmamba) for more details about performance evaluation.
 
-#### Software
 
-[More Information Needed]
+<br>
 
-## Citation [optional]
+# Technical Specifications 
 
-<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+## Model Architecture and Objective
 
-**BibTeX:**
+Falcon-Mamba-7B is a causal decoder-only model trained on a causal language modeling task (i.e., predict the next token).
 
-[More Information Needed]
+The model is based on the Mamba architecture ([Gu et al., 2023](https://arxiv.org/abs/2312.00752)).
 
-**APA:**
+| **Hyperparameter** | **Value** | **Comment**                            |
+|--------------------|-----------|----------------------------------------|
+| Layers             | 64        | Number of layers                       |
+| `d_model`          | 4096      | Hidden dimension                       |
+| `d_state`          | 16        | The SSM state dimension                |
+| Vocabulary         | 65024     | Vocabulary Size                        |
+| Sequence length    | 8192      | During the last training stages        |
 
-[More Information Needed]
+## Compute Infrastructure
 
-## Glossary [optional]
+### Hardware
 
-<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+Falcon-Mamba-7B was trained on AWS SageMaker, using on average 256 H100 80GB GPUs in 32 p5 instances. 
 
-[More Information Needed]
+### Software
 
-## More Information [optional]
+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.
 
-[More Information Needed]
+<br>
 
-## Model Card Authors [optional]
+# Citation
 
-[More Information Needed]
-
-## Model Card Contact
-
-[More Information Needed]
+*Paper coming soon* 😊. In the meanwhile, you can use the following information to cite: 
+```
+@article{falconmamba,
+  title={Falcon Mamba: The First Competitive Attention-free 7B Language Model},
+  author={Zuo, Jingwei and Velikanov, Maksim and Rhaiem, Dhia Eddine and Chahed, Ilyas and Belkada, Younes and Kunsch, Guillaume and Hacid, Hakim},
+  year={2024}
+}
+```