This repository contains the Falcon3-Mamba-7B-Instruct. It achieves, compared to similar SSM-based models of the same size, state of art results (at release's time) on reasoning, language understanding, instruction following, code and mathematics tasks. Falcon3-Mamba-7B-Instruct supports a context length up to 32K and was mainly trained on english corpus.

Model Details

Architecture (same as Falcon-Mamba-7b)
- Mamba1 based causal decoder only architecture trained on a causal language modeling task (i.e., predict the next token).
- 64 decoder blocks
- width: 4096
- state_size: 16
- 32k context length
- 65k vocab size
Continue Pretrained from Falcon-Mamba-7b, with another 1500 Gigatokens of data consisting of web, code, STEM and high quality data.
Postrained on 1.2 million samples of STEM, conversations, code, and safety.
Developed by Technology Innovation Institute
License: TII Falcon-LLM License 2.0
Model Release Date: December 2024

Getting started

Click to expand

python

from transformers import AutoTokenizer, AutoModelForCausalLM


from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "tiiuae/Falcon3-Mamba-7B-Instruct"

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

prompt = "How many hours in one day?"
messages = [
    {"role": "system", "content": "You are a helpful friendly assistant Falcon3 from TII, try to follow instructions as much as possible."},
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=1024
)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]

response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(response)

Benchmarks

We report in the following table our internal pipeline benchmarks. For the benchmarks marked by star, we normalize the results with HuggingFace score normalization:

Category	Benchmark	Zamba2-7B-instruct	Jamba-1.5-Mini	Llama-3.1-8B-Instruct	Falcon3-Mamba-7B-Instruct
General	MMLU (5-shot)	30.6	68.7	55.9	65.3
MMLU-PRO (5-shot)*	32.4	31.6	21.8	26.3
IFEval	69.9	65.7	78.8	71.7
Math	GSM8K (5-shot)	0	74.9	19.2	65.2
MATH Lvl-5 (4-shot)	13.6	6.9	10.4	27.3
Reasoning	Arc Challenge (25-shot)	54	54.3	46.6	53.7
GPQA (0-shot)*	10.3	11.1	6.2	7.2
MUSR (0-shot)*	8.2	12.2	38.6	8.3
BBH (3-shot)*	33.3	35.3	43.7	25.2
CommonSense Understanding	PIQA (0-shot)	75.6	82.3	78.9	80.9
SciQ (0-shot)	29.2	94.9	80.2	93.6
OpenbookQA (0-shot)	45.6	45.8	46.2	47.2

Useful links

View our release blogpost.
Feel free to join our discord server if you have any questions or to interact with our researchers and developers.

Citation

If the Falcon3 family of models were helpful to your work, feel free to give us a cite.

@misc{Falcon3,
    title = {The Falcon 3 Family of Open Models},
    author = {Falcon-LLM Team},
    month = {December},
    year = {2024}
}