|
|
--- |
|
|
library_name: transformers |
|
|
tags: |
|
|
- falcon-h1 |
|
|
license: other |
|
|
license_name: falcon-llm-license |
|
|
license_link: https://falconllm.tii.ae/falcon-terms-and-conditions.html |
|
|
base_model: tiiuae/Falcon-H1-34B-Instruct |
|
|
inference: true |
|
|
--- |
|
|
|
|
|
<img src="https://huggingface.co/datasets/tiiuae/documentation-images/resolve/main/falcon_mamba/falcon-h1-logo.png" alt="drawing" width="800"/> |
|
|
|
|
|
# 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:** Hybrid Transformers + Mamba architecture |
|
|
- **Language(s) (NLP):** English, Multilingual |
|
|
- **License:** Falcon-LLM License |
|
|
|
|
|
# Training details |
|
|
|
|
|
For more details about the training protocol of this model, please refer to the [Falcon-H1 technical blogpost](https://falcon-lm.github.io/blog/falcon-h1/). |
|
|
|
|
|
# Usage |
|
|
|
|
|
Currently to use this model you can either rely on Hugging Face `transformers`, `vLLM` or our custom fork of `llama.cpp` library. |
|
|
|
|
|
## Inference |
|
|
|
|
|
Make sure to install the latest version of `transformers` or `vllm`, eventually install these packages from source: |
|
|
|
|
|
```bash |
|
|
pip install git+https://github.com/huggingface/transformers.git |
|
|
``` |
|
|
|
|
|
Refer to [the official vLLM documentation for more details on building vLLM from source](https://docs.vllm.ai/en/latest/getting_started/installation/gpu.html#build-wheel-from-source). |
|
|
|
|
|
### 🤗 transformers |
|
|
|
|
|
Refer to the snippet below to run H1 models using 🤗 transformers: |
|
|
|
|
|
```python |
|
|
import torch |
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer |
|
|
|
|
|
model_id = "tiiuae/Falcon-H1-1B-Base" |
|
|
|
|
|
model = AutoModelForCausalLM.from_pretrained( |
|
|
model_id, |
|
|
torch_dtype=torch.bfloat16, |
|
|
device_map="auto" |
|
|
) |
|
|
|
|
|
# Perform text generation |
|
|
``` |
|
|
|
|
|
### vLLM |
|
|
|
|
|
For vLLM, simply start a server by executing the command below: |
|
|
|
|
|
``` |
|
|
# pip install vllm |
|
|
vllm serve tiiuae/Falcon-H1-1B-Instruct --tensor-parallel-size 2 --data-parallel-size 1 |
|
|
``` |
|
|
|
|
|
### 🦙 llama.cpp |
|
|
|
|
|
While we are working on integrating our architecture directly into `llama.cpp` library, you can install our fork of the library and use it directly: https://github.com/tiiuae/llama.cpp-Falcon-H1 |
|
|
Use the same installing guidelines as `llama.cpp`. |
|
|
|
|
|
# Evaluation |
|
|
|
|
|
Falcon-H1 series perform very well on a variety of tasks, including reasoning tasks. |
|
|
|
|
|
| Tasks | Falcon-H1-34B | Qwen3-32B | Qwen2.5-72B | Qwen2.5-32B | Gemma3-27B | Llama3.3-70B | Llama4-scout | |
|
|
| --- | --- | --- | --- | --- | --- | --- | --- | |
|
|
| **General** | | | | | | | |
|
|
| BBH | 70.68 | 62.47 | **72.52** | 68.72 | 67.28 | 69.15 | 64.9 | |
|
|
| ARC-C | 61.01 | 48.98 | 46.59 | 44.54 | 54.52 | **63.65** | 56.14 | |
|
|
| TruthfulQA | 65.27 | 58.58 | 69.8 | **70.28** | 64.26 | 66.15 | 62.74 | |
|
|
| HellaSwag | **81.94** | 68.89 | 68.79 | 73.95 | 57.25 | 70.24 | 65.03 | |
|
|
| MMLU | 84.05 | 80.89 | **84.42** | 82.8 | 78.01 | 82.08 | 80.4 | |
|
|
| **Math** | | | | | | | |
|
|
| GSM8k | 83.62 | 88.78 | 82.26 | 78.47 | 90.37 | **93.71** | 90.37 | |
|
|
| MATH-500 | 83.8 | 82.0 | 83.6 | 82.2 | **90.0** | 70.6 | 83.2 | |
|
|
| AMC-23 | 69.38 | 67.34 | 67.34 | 68.75 | **77.81** | 39.38 | 69.06 | |
|
|
| AIME-24 | 23.75 | 27.71 | 17.29 | 17.92 | 27.5 | 12.92 | **27.92** | |
|
|
| AIME-25 | 16.67 | 19.79 | 15.21 | 11.46 | **22.71** | 1.25 | 8.96 | |
|
|
| **Science** | | | | | | | |
|
|
| GPQA | **41.53** | 30.2 | 37.67 | 34.31 | 36.49 | 31.99 | 31.8 | |
|
|
| GPQA_Diamond | 49.66 | 49.49 | 44.95 | 40.74 | 47.47 | 42.09 | **51.18** | |
|
|
| MMLU-Pro | **58.73** | 54.68 | 56.35 | 56.63 | 47.81 | 53.29 | 55.58 | |
|
|
| MMLU-stem | **83.57** | 81.64 | 82.59 | 82.37 | 73.55 | 74.88 | 75.2 | |
|
|
| **Code** | | | | | | | |
|
|
| HumanEval | 87.2 | **90.85** | 87.2 | 90.24 | 86.59 | 83.53 | 85.4 | |
|
|
| HumanEval+ | 81.71 | **85.37** | 80.49 | 82.32 | 78.05 | 79.87 | 78.7 | |
|
|
| MBPP | 83.86 | 86.24 | **89.68** | 87.83 | 88.36 | 88.09 | 81.5 | |
|
|
| MBPP+ | 71.43 | 71.96 | **75.4** | 74.07 | 74.07 | 73.81 | 64.8 | |
|
|
| LiveCodeBench | 49.71 | 45.01 | **54.6** | 49.12 | 39.53 | 40.31 | 40.12 | |
|
|
| CRUXEval | 73.07 | **78.45** | 75.63 | 73.5 | 74.82 | 69.53 | 68.32 | |
|
|
| **Instruction Following** | | | | | | | |
|
|
| IFEval | 89.37 | 86.97 | 86.35 | 81.79 | 83.19 | **89.94** | 86.32 | |
|
|
| Alpaca-Eval | 48.32 | **64.21** | 49.29 | 39.26 | 56.16 | 38.27 | 36.26 | |
|
|
| MTBench | **9.2** | 9.05 | 9.16 | 9.09 | 8.75 | 8.98 | 8.98 | |
|
|
| LiveBench | 46.26 | **63.05** | 54.03 | 52.92 | 55.41 | 53.11 | 54.21 | |
|
|
|
|
|
You can check more in detail on our [our release blogpost](https://falcon-lm.github.io/blog/falcon-h1/), detailed benchmarks. |
|
|
|
|
|
# Useful links |
|
|
|
|
|
- View [our release blogpost](https://falcon-lm.github.io/blog/falcon-h1/). |
|
|
- Feel free to join [our discord server](https://discord.gg/trwMYP9PYm) if you have any questions or to interact with our researchers and developers. |
|
|
|
|
|
# Citation |
|
|
|
|
|
If the Falcon-H1 family of models were helpful to your work, feel free to give us a cite. |
|
|
|
|
|
``` |
|
|
@misc{tiifalconh1, |
|
|
title = {Falcon-H1: A Family of Hybrid-Head Language Models Redefining Efficiency and Performance}, |
|
|
url = {https://falcon-lm.github.io/blog/falcon-h1}, |
|
|
author = {Falcon-LLM Team}, |
|
|
month = {May}, |
|
|
year = {2025} |
|
|
} |
|
|
``` |
