<div align="center"> <img src="https://huggingface.co/datasets/tiiuae/documentation-images/resolve/main/general/falco3-logo.png" alt="drawing" width="500"/> </div>

Falcon3-10B-Instruct

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

This repository contains the Falcon3-10B-Instruct. It achieves state-of-the-art results (at the time of release) on reasoning, language understanding, instruction following, code and mathematics tasks. Falcon3-10B-Instruct supports 4 languages (English, French, Spanish, Portuguese) and a context length of up to 32K.

Model Details

• Architecture
  • Transformer-based causal decoder-only architecture
  • 40 decoder blocks
  • Grouped Query Attention (GQA) for faster inference: 12 query heads and 4 key-value heads
  • Wider head dimension: 256
  • High RoPE value to support long context understanding: 1000042
  • Uses SwiGLu and RMSNorm
  • 32K context length
  • 131K vocab size
• Depth up-scaled from Falcon3-7B-Base with 2 Teratokens of datasets comprising of web, code, STEM, high quality and mutlilingual data using 1024 H100 GPU chips
• Posttrained on 1.2 million samples of STEM, conversational, code, safety and function call data
• Supports EN, FR, ES, PT
• Developed by Technology Innovation Institute
• License: TII Falcon-LLM License 2.0
• Model Release Date: December 2024

Getting started

<details> <summary> Click to expand </summary>

from transformers import AutoTokenizer, AutoModelForCausalLM


from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "tiiuae/Falcon3-10B-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)

</details> <br>

Benchmarks

We report the official HuggingFace leaderboard normalized evaluations Open LLM Leaderboard Evaluation Results in the following table.

<table border="1" style="width: 100%; text-align: center; border-collapse: collapse;"> <colgroup> <col style="width: 10%;"> <col style="width: 7%;"> <col style="width: 7%;"> <col style="width: 7%;"> <col style="background-color: rgba(80, 15, 213, 0.5); width: 7%;"> </colgroup> <thead> <tr> <th>Benchmark</th> <th>Yi-1.5-9B-Chat</th> <th>Mistral-Nemo-Instruct-2407 (12B)</th> <th>Gemma-2-9b-it</th> <th>Falcon3-10B-Instruct</th> </tr> </thead> <tbody> <tr> <td>IFEval</td> <td>60.46</td> <td>63.80</td> <td>74.36</td> <td><b>78.17</b></td> </tr> <tr> <td>BBH (3-shot)</td> <td>36.95</td> <td>29.68</td> <td>42.14</td> <td><b>44.82</b></td> </tr> <tr> <td>MATH Lvl-5 (4-shot)</td> <td>12.76</td> <td>6.50</td> <td>0.23</td> <td><b>25.91</b></td> </tr> <tr> <td>GPQA (0-shot)</td> <td>11.30</td> <td>5.37</td> <td><b>14.77</b></td> <td>10.51</td> </tr> <tr> <td>MUSR (0-shot)</td> <td>12.84</td> <td>8.48</td> <td>9.74</td> <td><b>13.61</b></td> </tr> <tr> <td>MMLU-PRO (5-shot)</td> <td>33.06</td> <td>27.97</td> <td>31.95</td> <td><b>38.10</b></td> </tr> </tbody> </table>

Also, we report in the following table our internal pipeline benchmarks.

• We use lm-evaluation harness.
• We report raw scores obtained by applying chat template and fewshot_as_multiturn.
• We use same batch-size across all models.

<table border="1" style="width: 100%; text-align: center; border-collapse: collapse;"> <colgroup> <col style="width: 10%;"> <col style="width: 10%;"> <col style="width: 7%;"> <col style="width: 7%;"> <col style="background-color: rgba(80, 15, 213, 0.5); width: 7%;"> </colgroup> <thead> <tr> <th>Category</th> <th>Benchmark</th> <th>Yi-1.5-9B-Chat</th> <th>Mistral-Nemo-Instruct-2407 (12B)</th> <th>Falcon3-10B-Instruct</th> </tr> </thead> <tbody> <tr> <td rowspan="3">General</td> <td>MMLU (5-shot)</td> <td>68.8</td> <td>66.0</td> <td><b>73.9</b></td> </tr> <tr> <td>MMLU-PRO (5-shot)</td> <td>38.8</td> <td>34.3</td> <td><b>44</b></td> </tr> <tr> <td>IFEval</td> <td>57.8</td> <td>63.4</td> <td><b>78</b></td> </tr> <tr> <td rowspan="3">Math</td> <td>GSM8K (5-shot)</td> <td>77.1</td> <td>77.6</td> <td><b>84.9</b></td> </tr> <tr> <td>GSM8K (8-shot, COT)</td> <td>76</td> <td>80.4</td> <td><b>84.6</b></td> </tr> <tr> <td>MATH Lvl-5 (4-shot)</td> <td>3.3</td> <td>5.9</td> <td><b>22.1</b></td> </tr> <tr> <td rowspan="5">Reasoning</td> <td>Arc Challenge (25-shot)</td> <td>58.3</td> <td>63.4</td> <td><b>66.2</b></td> </tr> <tr> <td>GPQA (0-shot)</td> <td><b>35.6</b></td> <td>33.2</td> <td>33.5</td> </tr> <tr> <td>GPQA (0-shot, COT)</td> <td>16</td> <td>12.7</td> <td><b>32.6</b></td> </tr> <tr> <td>MUSR (0-shot)</td> <td><b>41.9</b></td> <td>38.1</td> <td>41.1</td> </tr> <tr> <td>BBH (3-shot)</td> <td>50.6</td> <td>47.5</td> <td><b>58.4</b></td> </tr> <tr> <td rowspan="4">CommonSense Understanding</td> <td>PIQA (0-shot)</td> <td>76.4</td> <td>78.2</td> <td><b>78.4</b></td> </tr> <tr> <td>SciQ (0-shot)</td> <td>61.7</td> <td>76.4</td> <td><b>90.4</b></td> </tr> <tr> <td>Winogrande (0-shot)</td> <td>-</td> <td>-</td> <td>71</td> </tr> <tr> <td>OpenbookQA (0-shot)</td> <td>43.2</td> <td>47.4</td> <td><b>48.2</b></td> </tr> <tr> <td rowspan="2">Instructions following</td> <td>MT-Bench (avg)</td> <td>8.3</td> <td><b>8.6</b></td> <td>8.2</td> </tr> <tr> <td>Alpaca (WC)</td> <td>25.8</td> <td><b>45.4</b></td> <td>24.7</td> </tr> <tr> <td>Tool use</td> <td>BFCL AST (avg)</td> <td>48.4</td> <td>74.2</td> <td><b>90.5</b></td> </tr> <tr> <td rowspan="2">Code</td> <td>EvalPlus (0-shot) (avg)</td> <td>69.4</td> <td>58.9</td> <td><b>74.7</b></td> </tr> <tr> <td>Multipl-E (0-shot) (avg)</td> <td>-</td> <td>34.5</td> <td><b>45.8</b></td> </tr> </tbody> </table>

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.

Technical Report

Coming soon....

Citation

If Falcon3 family were helpful in 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}
}

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	35.19
IFEval (0-Shot)	78.17
BBH (3-Shot)	44.82
MATH Lvl 5 (4-Shot)	25.91
GPQA (0-shot)	10.51
MuSR (0-shot)	13.61
MMLU-PRO (5-shot)	38.10