File size: 3,459 Bytes
1a56f20 6ab5791 1a56f20 |
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 |
---
license: llama2
datasets:
- tiiuae/falcon-refinedweb
- EleutherAI/pile
- meta-math/MetaMathQA
language:
- en
library_name: transformers
---
# Saily 220B
<img src="https://i.ibb.co/rG8S6cF/Saily-220-B.png" style="width: 100%; height: auto;"/>
---
## Announcements
**1.** <b>Date: </b>17th December, 2023
Releasing v1. Saily_220B is a powerful AI model built on top of Llama2-70B merges.
We created 10 fine-tuned **Llama2 70B** models. The models were fine-tuned on a part of Refined-Web Dataset (common for all)
and individually the models were finetuned on niche specific datasets:
- Code
- Humor
- Maths
- Logical Understanding
- Physics
- Reasoning
- Psychology
- Roleplay
We created 4 linear merges while keeping **Logical-Understanding** and **Reasoning** models constant in all linear merges.
and then finally we created a passthrough merge between the models.
Public Datasets used:
1. [RefinedWeb](https://hf.co/datasets/tiiuae/falcon-refinedweb) (part of it)
2. Pile (part of it)
3. [MetaMathQA](https://hf.co/datasets/meta-math/MetaMathQA)
4. Unnatural Code (Javascript, Python, C++)
### How did we create the private dataset?
We recorded many internal brain-storming sessions where we just talked about random things.
We also invited many experts from different fields:
- Mathematicians
- Developers
- Bio-Engineers
- Authors
- Psychologists
- and others...
We talked about different things with them and recorded the sessions and then transcribed the audio to create the datasets.
---
### Please don't refer to the config.json in the files, it isn't accurate. You can run:
```python
from transformers import AutoModelForCausalLM as amclm
model = amclm.from_pretrained("deepnight-research/saily_220b",
device_map="auto")
# print(model.config)
model.config
```
to check out the model's configuration.
---
### Try it:
You definitely need GPUs here (that goes without saying)
* We have tried it on **4 x A100 80GB** and **2 x A100 80GB**.
* You will have to load the model in **4bit** to fit on **2 x A100 (80GB)**.
```python
from transformers import AutoModelForCausalLM as amclm
from transformers import AutoTokenizer
model_name = "deepnight-research/saily_220b"
model = amclm.from_pretrained(model_name, device_map="auto")
# To load in 8Bit, make sure you have bitsandbytes installed.
# model = amclm.from_pretrained(model_name,
# device_map="auto",
# load_in_8bit=True
# )
# Float16
# import torch
# model = amclm.from_pretrained(model_name,
# device_map="auto",
# torch_dtype=torch.float16
# )
tokenizer = AutoTokenier.from_pretrained(model_name)
input_ids = tokenizer.encode("[INST]\nWrite a poem about cats\n[/INST]\n\n", return_tensors="pt")
output = model.generate(input_ids, max_length=128,
temperature=0.7,
repetition_penalty=1.1,
top_p=0.7, top_k=50
)
output_text = tokenizer.decode(output[0], skip_special_tokens=True)
```
We recommend following **Alpaca Prompt Format**, and if you're trying it out in Text-Generation-WebUI, please use **INSTRUCT** or **CHAT-INSTRUCT** mode.
---
## Limitations and Bias
As with all language models, Saily_220B may generate incorrect or biased content. It's important to keep this in mind when using the model.
---
## Wanna Talk?
Reach out to us at [[email protected]](mailto:[email protected]) or [[email protected]](mailto:[email protected]) |