Model card for FalconMamba Instruct model

Table of Contents

1. TL;DR
2. Model Details
3. Usage
4. Training Details
5. Evaluation

TL;DR

Model Details

Model Description

• Developed by: https://www.tii.ae
• Model type: Causal decoder-only
• Architecture: Mamba
• Language(s) (NLP): Mainly English
• License: TII Falcon-Mamba License 2.0

Usage

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):

Using the Pytorch model

Running the model on a CPU

Click to expand

```python from transformers import AutoTokenizer, AutoModelForCausalLM tokenizer = AutoTokenizer.from_pretrained("tiiuae/falcon-mamba-7b-instruct") model = AutoModelForCausalLM.from_pretrained("tiiuae/falcon-mamba-7b-instruct") # We use the tokenizer's chat template to format each message - see https://huggingface.co/docs/transformers/main/en/chat_templating messages = [ {"role": "user", "content": "How many helicopters can a human eat in one sitting?"}, ] input_text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True) input_ids = tokenizer(input_text, return_tensors="pt").input_ids outputs = model.generate(input_ids, max_new_tokens=30) print(tokenizer.decode(outputs[0])) ```

Running the model on a GPU

Click to expand

```python # pip install accelerate from transformers import AutoTokenizer, AutoModelForCausalLM tokenizer = AutoTokenizer.from_pretrained("tiiuae/falcon-mamba-7b-instruct") model = AutoModelForCausalLM.from_pretrained("tiiuae/falcon-mamba-7b-instruct", device_map="auto") # We use the tokenizer's chat template to format each message - see https://huggingface.co/docs/transformers/main/en/chat_templating messages = [ {"role": "user", "content": "How many helicopters can a human eat in one sitting?"}, ] input_text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True) input_ids = tokenizer(input_text, return_tensors="pt").input_ids.to("cuda") outputs = model.generate(input_ids, max_new_tokens=30) print(tokenizer.decode(outputs[0])) ```

Running the model on a GPU using torch.compile

Click to expand

```python import torch from transformers import AutoTokenizer, AutoModelForCausalLM tokenizer = AutoTokenizer.from_pretrained("tiiuae/falcon-mamba-7b-instruct") model = AutoModelForCausalLM.from_pretrained("tiiuae/falcon-mamba-7b-instruct", torch_dtype=torch.bfloat16).to(0) model = torch.compile(model) # We use the tokenizer's chat template to format each message - see https://huggingface.co/docs/transformers/main/en/chat_templating messages = [ {"role": "user", "content": "How many helicopters can a human eat in one sitting?"}, ] input_text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True) input_ids = tokenizer(input_text, return_tensors="pt").input_ids.to("cuda") outputs = model.generate(input_ids, max_new_tokens=30) print(tokenizer.decode(outputs[0])) ```

Running the model on a GPU using different precisions

FP16

Click to expand

```python # pip install accelerate import torch from transformers import AutoTokenizer, AutoModelForCausalLM tokenizer = AutoTokenizer.from_pretrained("tiiuae/falcon-mamba-7b-instruct") model = AutoModelForCausalLM.from_pretrained("tiiuae/falcon-mamba-7b-instruct", device_map="auto", torch_dtype=torch.float16) # We use the tokenizer's chat template to format each message - see https://huggingface.co/docs/transformers/main/en/chat_templating messages = [ {"role": "user", "content": "How many helicopters can a human eat in one sitting?"}, ] input_text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True) input_ids = tokenizer(input_text, return_tensors="pt").input_ids.to("cuda") outputs = model.generate(input_ids, max_new_tokens=30) print(tokenizer.decode(outputs[0])) ```

4-bit

Click to expand

```python # pip install bitsandbytes accelerate from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig tokenizer = AutoTokenizer.from_pretrained("tiiuae/falcon-mamba-7b-instruct") model = AutoModelForCausalLM.from_pretrained("tiiuae/falcon-mamba-7b-instruct", device_map="auto", quantization_config=BitsAndBytesConfig(load_in_4bit=True)) # We use the tokenizer's chat template to format each message - see https://huggingface.co/docs/transformers/main/en/chat_templating messages = [ {"role": "user", "content": "How many helicopters can a human eat in one sitting?"}, ] input_text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True) input_ids = tokenizer(input_text, return_tensors="pt").input_ids.to("cuda") outputs = model.generate(input_ids, max_new_tokens=30) print(tokenizer.decode(outputs[0])) ```

Training Details

Training Data

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.

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.

The data was tokenized with the Falcon-7B/11B tokenizer.

After pre-training, the model has been further fine-tuned on instruction data.

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.

Training Hyperparameters

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

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.

Evaluation

Benchmarks

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.

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*	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*	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

Also, we evaluate our model on the benchmarks of the first leaderboard using lighteval.

model name	ARC	HellaSwag	MMLU	Winogrande	TruthfulQA	GSM8K	Average
Pure SSM models
FalconMamba-7B*	62.03	80.82	62.11	73.64	53.42	52.54	64.09
TRI-ML/mamba-7b-rw*	51.25	80.85	33.41	71.11	32.08	4.70	45.52
Hybrid SSM-attention models
recurrentgemma-9b**	52.00	80.40	60.50	73.60	38.60	42.60	57.95
Zyphra/Zamba-7B-v1*	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
gemma-7B	61.09	82.20	64.56	79.01	44.79	50.87	63.75

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.

Throughput

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:

pip install "causal-conv1d>=1.4.0" mamba-ssm

Refer to our FalconMamba blogpost for more details about performance evaluation.

Technical Specifications

Model Architecture and Objective

Falcon-Mamba-7B is a causal decoder-only model trained on a causal language modeling task (i.e., predict the next token).

The model is based on the Mamba architecture (Gu et al., 2023).

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

Compute Infrastructure

Hardware

Falcon-Mamba-7B was trained on AWS SageMaker, using on average 256 H100 80GB GPUs in 32 p5 instances.

Software

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.

Citation

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}
}