README.md

---
library_name: transformers
tags:
- bitnet
- falcon3
base_model: tiiuae/Falcon3-3B-Base
---

![image/png](https://cdn-uploads.huggingface.co/production/uploads/62441d1d9fdefb55a0b7d12c/c-tosr0FvMlKuKQTojx_6.png)


#  Table of Contents

0. [TL;DR](#TL;DR)
1. [Model Details](#model-details)
2. [Training Details](#training-details)
3. [Usage](#usage)
4. [Evaluation](#evaluation)
5. [Citation](#citation)


# TL;DR

# Model Details

## Model Description

- **Developed by:** [https://www.tii.ae](https://www.tii.ae)
- **Model type:** Causal decoder-only
- **Architecture:** Pure-transformer - 1.58bit version
- **Language(s) (NLP):** Mainly English
- **License:** TII Falcon License 2.0

# Training details

The model has been trained following the training strategies from the recent [1-bit LLM HF blogpost](https://huggingface.co/blog/1_58_llm_extreme_quantization) and [1-bit LLM paper](https://huggingface.co/papers/2402.17764).
For more details about the training protocol of this model, please refer to the Falcon-3 technical report, section *Compression*.


# Usage

Currently to use this model you can either rely on Hugging Face transformers library or [BitNet](https://github.com/microsoft/BitNet) library. You can also play with the model using the [falcon-1.58bit playground](https://huggingface.co/spaces/tiiuae/falcon3-1.58bit-playground) (only for the 7B instruct version).

## 🤗 transformers

```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "tiiuae/Falcon3-3B-Base-1.58bit"

model = AutoModelForCausalLM.from_pretrained(
  model_id,
  torch_dtype=torch.bfloat16,
).to("cuda")

# Perform text generation
```

## BitNet

```
git clone https://github.com/microsoft/BitNet && cd BitNet
pip install -r requirements.txt
python setup_env.py --hf-repo tiiuae/Falcon3-3B-Base-1.58bit -q i2_s
python run_inference.py -m models/Falcon3-3B-Base-1.58bit/ggml-model-i2_s.gguf -p "Hi how are you doing today?" -cnv
```

# Evaluation
We report in the following table our internal pipeline benchmarks:

**Note evaluation results are normalized score from v2 leaderboard tasks - reported results of original models in the blogpost are raw scores**

<table border="1" style="width: 100%; text-align: center; border-collapse: collapse;">
    <colgroup>
        <col style="width: 10%;">
        <col style="width: 10%;">
        <col style="background-color: rgba(80, 15, 213, 0.5); width: 7%;">
    </colgroup>
    <thead>
        <tr>
            <th>Benchmark</th>
            <th>Llama3-8B-1.58-100B-tokens</th>
            <th>Falcon3-7B-Instruct-1.58bit </th>
        </tr>
    </thead>
    <tbody>
        <tr>
            <td>IFEval</td>
            <td>17.91</td>
            <td><b>27.49</b></td>
        </tr>      
        <tr>
            <td>MUSR</td>
            <td><b>4.87</b></td>
            <td>4.64</td>
        </tr>
        <tr>
            <td>GPQA</td>
            <td><b>1.83<b></td>
            <td>0.00</td>
        </tr>
        <tr>
            <td>BBH</td>
            <td>5.36</td>
            <td><b>2.97</b></td>
        </tr>
        <tr>
            <td>MMLU-PRO</td>
            <td><b>2.78<b></td>
            <td><b>1.47</b></td>
        </tr>      
        <tr>
            <td>MATH</td>
            <td>0.26</td>
            <td><b>0.43</b></td>
        </tr>
        <tr>
            <td>Average</td>
            <td>5.5</td>
            <td><b>6.17</b></td>
        </tr>          
    </tbody>
</table>

# Citation

Coming soon ..