File size: 3,099 Bytes
dc660c4
ad3b4ce
 
 
 
 
 
 
 
 
 
 
 
dc660c4
 
ad3b4ce
a4f5e88
ad3b4ce
f085e09
ad3b4ce
 
 
 
 
 
 
 
 
f085e09
ad3b4ce
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e5b2e12
ad3b4ce
 
 
 
f085e09
ad3b4ce
 
 
 
 
 
 
 
 
f085e09
ad3b4ce
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f085e09
ad3b4ce
 
 
 
 
 
 
 
 
 
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
---
language: 
  - en
tags:
- llama2
- llama-2
- llama
- llama2 architecture
- litellama
datasets:
- Redpajama
metrics:
- MMLU
license: mit
---

# LiteLlama: Reduced-Scale Llama

In this series of repos, we present an open-source reproduction of Meta AI's [LLaMa 2](https://ai.meta.com/llama/). However, with significantly reduced model sizes, [LiteLlama-460M-1T](https://huggingface.co/ahxt/LiteLlama-460M-1T) has 460M parameters trained with 1T tokens.


## Dataset and Tokenization
We train our models on part of [RedPajama](https://www.together.xyz/blog/redpajama) dataset. We use the [GPT2Tokenizer](https://huggingface.co/docs/transformers/v4.31.0/en/model_doc/gpt2#transformers.GPT2Tokenizer) to tokenize the text.

## Training Details

The model was trained with ~1T tokens (0.98T). num of tokens = steps*length*batch_size=499679*1024*192=98240888832≈0.98T.

The training curve is at this [WandB project](https://wandb.ai/ahxt/llama2_xs_460M_training_loss/reports/reduced_train_loss-23-09-05-20-25-43---Vmlldzo1MzIwNDUx?accessToken=x2ch3n30jo77p1x8y7q9js4h4d8zpjtz1tzot4xxullyefixp4jwt7au2q37k2q6).

### Using with HuggingFace Transformers
The experimental checkpoints can be directly loaded by [Transformers](https://huggingface.co/transformers/) library. The following code snippet shows how to load the our experimental model and generate text with it. 


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

model_path = 'ahxt/LiteLlama-460M-1T'

model = AutoModelForCausalLM.from_pretrained(model_path)
tokenizer = AutoTokenizer.from_pretrained(model_path)
model.eval()

prompt = 'Q: What is the largest bird?\nA:'
input_ids = tokenizer(prompt, return_tensors="pt").input_ids
tokens = model.generate(input_ids, max_length=20)
print( tokenizer.decode(tokens[0].tolist(), skip_special_tokens=True) )
# Q: What is the largest bird?\nA: The largest bird is a black-headed gull.
```

## Evaluation

### We evaluate our models on the MMLU task.

| Models | #parameters |zero-shot |  5-shot |
| --- | --- | --- | --- |
| llama                       | 7B    | 28.46 | 35.05 |
| openllama                   | 3B    | 24.90 | 26.71 |
|TinyLlama-1.1B-step-50K-105b | 1.1B  | 19.00 | 26.53 |
| LiteLlama-460M-1T           | 0.46B | 21.13 | 26.39 |


### [Open LLM Leaderboard Evaluation Results](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard)
Detailed results can be found [here](https://huggingface.co/datasets/open-llm-leaderboard/details_ahxt__llama2_xs_460M_experimental)

| Metric                | Value                     |
|-----------------------|---------------------------|
| Avg.                  | 26.65   |
| ARC (25-shot)         | 24.91          |
| HellaSwag (10-shot)   | 38.47    |
| MMLU (5-shot)         | 26.17         |
| TruthfulQA (0-shot)   | 41.59   |
| Winogrande (5-shot)   | 49.88   |
| GSM8K (5-shot)        | 0.0        |
| DROP (3-shot)         | 5.51         |




## Contact
This model is developed by [Xiaotian Han](https://ahxt.github.io/) from Texas A&M University and released under MIT License.