File size: 9,232 Bytes
0665b94 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 |
---
datasets:
- tiiuae/falcon-refinedweb
---
# 🚀 Falcon-11B
**Falcon-11B is a 11B parameters causal decoder-only model built by [TII](https://www.tii.ae) and trained over 5,000B tokens of [RefinedWeb](https://huggingface.co/datasets/tiiuae/falcon-refinedweb) enhanced with curated corpora. The model is made available under the Apache 2.0 license.**
*Paper coming soon 😊.*
🤗 To get started with Falcon (inference, finetuning, quantization, etc.), we recommend reading [this great blogpost fron HF](https://huggingface.co/blog/falcon)!
⚠️ **This is a raw, pretrained model, which should be further finetuned for most usecases.** If you are looking for a version better suited to taking generic instructions in a chat format, we recommend taking a look at [Falcon-11B-Chat](https://huggingface.co/tiiuae/falcon-11B-chat).
```python
from transformers import AutoTokenizer, AutoModelForCausalLM
import transformers
import torch
model = "tiiuae/falcon-11B"
tokenizer = AutoTokenizer.from_pretrained(model)
pipeline = transformers.pipeline(
"text-generation",
model=model,
tokenizer=tokenizer,
torch_dtype=torch.bfloat16,
trust_remote_code=True,
device_map="auto",
)
sequences = pipeline(
"Girafatron is obsessed with giraffes, the most glorious animal on the face of this Earth. Giraftron believes all other animals are irrelevant when compared to the glorious majesty of the giraffe.\nDaniel: Hello, Girafatron!\nGirafatron:",
max_length=200,
do_sample=True,
top_k=10,
num_return_sequences=1,
eos_token_id=tokenizer.eos_token_id,
)
for seq in sequences:
print(f"Result: {seq['generated_text']}")
```
💥 **Falcon LLMs require PyTorch 2.0 for use with `transformers`!**
For fast inference with Falcon, check-out [Text Generation Inference](https://github.com/huggingface/text-generation-inference)! Read more in this [blogpost]((https://huggingface.co/blog/falcon).
# Model Card for Falcon-11B
## Model Details
### Model Description
- **Developed by:** [https://www.tii.ae](https://www.tii.ae);
- **Model type:** Causal decoder-only;
- **Language(s) (NLP):** English, German, Spanish, French, Italian, Portuguese, Polish, Dutch, Romanian, Czech, Swedish
- **License:** TII Falcon License 2.0
### Model Source
- **Paper:** *coming soon*.
## Uses
### Direct Use
Research on large language models; as a foundation for further specialization and finetuning for specific usecases (e.g., summarization, text generation, chatbot, etc.)
### Out-of-Scope Use
Production use without adequate assessment of risks and mitigation; any use cases which may be considered irresponsible or harmful.
## Bias, Risks, and Limitations
Falcon-11B is trained mostly on English, but also German, Spanish, French, Italian, Portuguese, Polish, Dutch, Romanian, Czech, Swedish. It will not generalize appropriately to other languages. Furthermore, as it is trained on a large-scale corpora representative of the web, it will carry the stereotypes and biases commonly encountered online.
### Recommendations
We recommend users of Falcon-11B to consider finetuning it for the specific set of tasks of interest, and for guardrails and appropriate precautions to be taken for any production use.
## How to Get Started with the Model
```python
from transformers import AutoTokenizer, AutoModelForCausalLM
import transformers
import torch
model = "tiiuae/falcon-11B"
tokenizer = AutoTokenizer.from_pretrained(model)
pipeline = transformers.pipeline(
"text-generation",
model=model,
tokenizer=tokenizer,
torch_dtype=torch.bfloat16,
trust_remote_code=True,
device_map="auto",
)
sequences = pipeline(
"Girafatron is obsessed with giraffes, the most glorious animal on the face of this Earth. Giraftron believes all other animals are irrelevant when compared to the glorious majesty of the giraffe.\nDaniel: Hello, Girafatron!\nGirafatron:",
max_length=200,
do_sample=True,
top_k=10,
num_return_sequences=1,
eos_token_id=tokenizer.eos_token_id,
)
for seq in sequences:
print(f"Result: {seq['generated_text']}")
```
## Training Details
### Training Data
Falcon-11B was trained over 5,000B tokens of [RefinedWeb](https://huggingface.co/datasets/tiiuae/falcon-refinedweb), a high-quality filtered and deduplicated web dataset which we enhanced with curated corpora. It followed a 4 stage training strategy. The first three stages being focused on increasing the context length, from to 2048 to 4096 and finally to 8192 tokens. The last stage aimed to further enhance performance using only high quality data.
Overall, the data sources included RefinedWeb-English, Refined Web-Europe (en, de, es, fr, it, pt, pl, nl, ro, sv, cs), high quality technical data, code data, and conversational data extracted from public sources.
| Technical | 2% | 20B | arXiv, PubMed, USPTO, etc. |
The training stages were as follows:
| **Stage** | **Context length** | ** Tokens** |
|--------------|-----------------|-------------|
| Stage 1 | 2048 | 4500B |
| Stage 2 | 4096 | 250B |
| Stage 3 | 8192 | 250B |
| Stage 4 | 8192 | 240B |
The data was tokenized with the Falcon-[7B](https://huggingface.co/tiiuae/falcon-7b)/[11B](https://huggingface.co/tiiuae/falcon-11B) tokenizer.
### Training Procedure
Falcon-11B was trained on 1024 A100 40GB GPUs, using a 3D parallelism strategy (TP=8, PP=1, DP=128) combined with ZeRO.
#### Training Hyperparameters
| **Hyperparameter** | **Value** | **Comment** |
|--------------------|------------|-------------------------------------------|
| Precision | `bfloat16` | |
| Optimizer | AdamW | |
| Max learning rate | 3.7e-4 | Following a linear warm=up, then cosine decay to 1.89e-5 across 4500 B tokens. |
| Weight decay | 1e-1 | |
| Z-loss | 1e-4 | |
| Batch size | Variable | Batch size was gradually increased duringthe training |
#### Speeds, Sizes, Times
The model training took roughly two months.
## Evaluation
|English Benchmark | **Value** | **Comment** |
|--------------------|------------|-------------------------------------------|
| HellaSwag-10shots | 82.91 | |
| Winogrande-5shots | 78.30 | |
| ARC-Challenge-25shots | 59.73 | |
| TruthfulQA-0shot | 52.56 | |
| MMLU-5shots | 58.37 | |
| GSM8k-5shots | 53.83 | |
| ARC-Challenge-0shot | 50.17 | |
| ARC-Easy-0shot | 77.78 | |
We thank the leaderboard team from HuggingFace for providing an official evaluation of our model on the leaderboard tasks.
## Technical Specifications
### Model Architecture and Objective
Falcon-11B is a causal decoder-only model trained on a causal language modeling task (i.e., predict the next token).
The architecture is broadly adapted from the GPT-3 paper ([Brown et al., 2020](https://arxiv.org/abs/2005.14165)), with the following differences:
* **Positionnal embeddings:** rotary ([Su et al., 2021](https://arxiv.org/abs/2104.09864));
* **Attention:** multiquery ([Shazeer et al., 2019](https://arxiv.org/abs/1911.02150)) and FlashAttention ([Dao et al., 2022](https://arxiv.org/abs/2205.14135));
* **Decoder-block:** parallel attention/MLP with a two layer norms.
For multiquery, we are using an internal variant which uses independent key and values per tensor parallel degree.
| **Hyperparameter** | **Value** | **Comment** |
|--------------------|-----------|----------------------------------------|
| Layers | 60 | |
| `d_model` | 4096 | |
| `head_dim` | 128 | |
| Vocabulary | 65024 | |
| Sequence length | 8192 | During stages 3 and 4 |
### Compute Infrastructure
#### Hardware
Falcon-11B was trained on AWS SageMaker, using on average 1024 A100 40GB GPUs in 128 p4d instances.
#### Software
Falcon-11B was trained a custom distributed training codebase, Gigatron. It uses a 3D parallelism approach combined with ZeRO and high-performance Triton kernels (FlashAttention2, etc.)
## Citation
*Paper coming soon* 😊.
## License
Falcon2 11B is licenced under TII Falcon License 2.0, the permissive Apache 2.0-based software license which includes an acceptable use policy that promotes the responsible use of AI.
## Contact
[email protected] |