metadata

language:
  - en
  - fr
  - es
  - pt
tags:
  - falcon3

Falcon3-7B-Base

Falcon3 family of Open Foundation Models is a set of pretrained and instruct LLMs ranging from 1B to 10B.

This repository contains the Falcon3-3B-Base. It achieves strong results on reasoning, language understanding, instruction following, code and mathematics tasks. Falcon3-3B-Base supports 4 languages (english, french, spanish, portuguese) and a context length up to 8K. Falcon3-3B-Base pruned (depth + width) from Falcon3-7B-Base, was effeciently trained on only 100 GT using a knowledge distillation objective.

⚠️ This is a raw, pretrained model, which should be further finetuned for most usecases.

Model Details

Architecture
- Transformer based causal decoder only architecture
- 22 decoder blocks
- Grouped query attention (GQA) for faster inference: 12 query heads and 4 KV heads
- Wider head dimension: 256
- High RoPE value to support long context understanding: 1000042
- 8k context length
- 131k vocab size
Pruned and Healed from Falcon3-7B-Base on only 100 Gigatokens of datasets comprising of web, code, STEM, high quality and mutlilingual data using 2048 H100 GPU chips
Supports EN, FR, ES, PT
Developed by Technology Innovation Institute
License: TII Falcon-LLM License 2.0
Model Release Date: December 2024

Getting started

Click to expand

import torch
from transformers import pipeline

pipe = pipeline(
    "text-generation", 
    model="tiiuae/Falcon3-3B-Base", 
    torch_dtype=torch.bfloat16, 
    device_map="auto"
)
response = pipe("Question: How many hours in one day? Answer: ")
print(response[0]['generated_text'])

Benchmarks

We report in the following table our internal pipeline benchmarks:

    </tr>
    <tr>
        <td rowspan="2">Math</td>
        <td>GSM8K (5-shot)</td>
        <td>26.68</td>
        <td>68.99</td>
        <td>25.7</td>
        <td>63.91</td>
    </tr>
    <tr>
        <td>MATH(4-shot)</td>
        <td>1.39</td>
        <td>8.43</td>
        <td>1.73</td>
        <td>9.38</td>
    </tr>
    <tr>
        <td rowspan="4">Reasoning</td>
        <td>Arc Challenge (25-shot)</td>
        <td>50.76</td>
        <td>55.54</td>
        <td>50.34</td>
        <td>54.86</td>
    </tr>
    <tr>
        <td>GPQA (0-shot)</td>
        <td>27.49</td>
        <td>27.53</td>
        <td>38.6</td>
        <td>31.15</td>
    </tr>
    <tr>
        <td>MUSR (0-shot)</td>
        <td>35.24</td>
        <td>43.03</td>
        <td>42.13</td>
        <td>37.5</td>
    </tr>
    <tr>
        <td>BBH (3-shot)</td>
        <td>38.59</td>
        <td>46.12</td>
        <td>40.85</td>
        <td>44.23</td>
    </tr>
    <tr>
        <td rowspan="4">CommonSense Understanding</td>
        <td>PIQA (0-shot)</td>
        <td>77.42</td>
        <td>78.89</td>
        <td>78.29</td>
        <td>75.62</td>
    </tr>
    <tr>
        <td>SciQ (0-shot)</td>
        <td>92.7</td>
        <td>95.6</td>
        <td>96.1</td>
        <td>93.1</td>
    </tr>
    <tr>
        <td>Winogrande (0-shot)</td>
        <td>69.69</td>
        <td>68.82</td>
        <td>68.35</td>
        <td>64.64</td>
    </tr>
    <tr>
        <td>OpenbookQA (0-shot)</td>
        <td>43.2</td>
        <td>42.2</td>
        <td>43</td>
        <td>39.4</td>
    </tr>
</tbody>

Category	Benchmark	Llama3.2-3B	Qwen2.5-3B	Minitron-4B	Falcon3-3B-Base
General	MMLU (5-shot)	56.1	65.6	58.6	55.5
	MMLU-PRO (5-shot)	24.9	31.99	26.21	28.77
	IFEval	12.83	27	22.81	27.67

Citation

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

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