Commit
•
962201e
1
Parent(s):
9832b62
release ORPO fine-tunes
Browse files
README.md
ADDED
@@ -0,0 +1,216 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: other
|
3 |
+
license_name: apple-sample-code-license
|
4 |
+
license_link: LICENSE
|
5 |
+
datasets:
|
6 |
+
- argilla/dpo-mix-7k
|
7 |
+
tags:
|
8 |
+
- apple
|
9 |
+
- OpenELM
|
10 |
+
---
|
11 |
+
|
12 |
+
# OpenELM
|
13 |
+
These are ORPO fine-tunes, done using the Argilla/dpo-mix-7k dataset:
|
14 |
+
- [270M fine-tune](https://huggingface.co/Trelis/OpenELM-270M-instruct-ORPO)
|
15 |
+
- [450M fine-tune](https://huggingface.co/Trelis/OpenELM-450M-instruct-ORPO)
|
16 |
+
|
17 |
+
## Performance notes
|
18 |
+
OpenELM models are quite weak.
|
19 |
+
- OpenELM 270M is uniquely small, but weak.
|
20 |
+
- OpenELM 450M improves a little over the 270M model, but remains weak on accuracy and hallucinates strongly.
|
21 |
+
- Qwen 1.5 0.5B is stronger than the OpenELM model.
|
22 |
+
- TinyLlama is stronger than OpenELM 1B.
|
23 |
+
- Models like Phi-3 are stronger than OpenELM 3B.
|
24 |
+
|
25 |
+
## Usage Notes
|
26 |
+
- Flash attention is not supported
|
27 |
+
- Making GGUFs is not [yet supported](https://github.com/ggerganov/llama.cpp/issues/6868)
|
28 |
+
|
29 |
+
## Inference
|
30 |
+
See [this Colab Notebook](https://colab.research.google.com/drive/1vFMRhHdPyUxbZAlRWwyl79NwnrSz_yQL?usp=sharing)
|
31 |
+
|
32 |
+
~~~
|
33 |
+
The original model card follows below.
|
34 |
+
~~~
|
35 |
+
|
36 |
+
*Sachin Mehta, Mohammad Hossein Sekhavat, Qingqing Cao, Maxwell Horton, Yanzi Jin, Chenfan Sun, Iman Mirzadeh, Mahyar Najibi, Dmitry Belenko, Peter Zatloukal, Mohammad Rastegari*
|
37 |
+
|
38 |
+
We introduce **OpenELM**, a family of **Open** **E**fficient **L**anguage **M**odels. OpenELM uses a layer-wise scaling strategy to efficiently allocate parameters within each layer of the transformer model, leading to enhanced accuracy. We pretrained OpenELM models using the [CoreNet](https://github.com/apple/corenet) library. We release both pretrained and instruction tuned models with 270M, 450M, 1.1B and 3B parameters.
|
39 |
+
|
40 |
+
Our pre-training dataset contains RefinedWeb, deduplicated PILE, a subset of RedPajama, and a subset of Dolma v1.6, totaling approximately 1.8 trillion tokens. Please check license agreements and terms of these datasets before using them.
|
41 |
+
|
42 |
+
|
43 |
+
|
44 |
+
## Usage
|
45 |
+
|
46 |
+
We have provided an example function to generate output from OpenELM models loaded via [HuggingFace Hub](https://huggingface.co/docs/hub/) in `generate_openelm.py`.
|
47 |
+
|
48 |
+
You can try the model by running the following command:
|
49 |
+
```
|
50 |
+
python generate_openelm.py --model apple/OpenELM-450M-Instruct --hf_access_token [HF_ACCESS_TOKEN] --prompt 'Once upon a time there was' --generate_kwargs repetition_penalty=1.2
|
51 |
+
```
|
52 |
+
Please refer to [this link](https://huggingface.co/docs/hub/security-tokens) to obtain your hugging face access token.
|
53 |
+
|
54 |
+
Additional arguments to the hugging face generate function can be passed via `generate_kwargs`. As an example, to speedup the inference, you can try [lookup token speculative generation](https://huggingface.co/docs/transformers/generation_strategies) by passing the `prompt_lookup_num_tokens` argument as follows:
|
55 |
+
```
|
56 |
+
python generate_openelm.py --model apple/OpenELM-450M-Instruct --hf_access_token [HF_ACCESS_TOKEN] --prompt 'Once upon a time there was' --generate_kwargs repetition_penalty=1.2 prompt_lookup_num_tokens=10
|
57 |
+
```
|
58 |
+
Alternatively, try model-wise speculative generation with an [assistive model](https://huggingface.co/blog/assisted-generation) by passing a smaller model through the `assistant_model` argument, for example:
|
59 |
+
```
|
60 |
+
python generate_openelm.py --model apple/OpenELM-450M-Instruct --hf_access_token [HF_ACCESS_TOKEN] --prompt 'Once upon a time there was' --generate_kwargs repetition_penalty=1.2 --assistant_model [SMALLER_MODEL]
|
61 |
+
```
|
62 |
+
|
63 |
+
## Main Results
|
64 |
+
|
65 |
+
### Zero-Shot
|
66 |
+
|
67 |
+
| **Model Size** | **ARC-c** | **ARC-e** | **BoolQ** | **HellaSwag** | **PIQA** | **SciQ** | **WinoGrande** | **Average** |
|
68 |
+
|-----------------------------------------------------------------------------|-----------|-----------|-----------|---------------|-----------|-----------|----------------|-------------|
|
69 |
+
| [OpenELM-270M](https://huggingface.co/apple/OpenELM-270M) | 26.45 | 45.08 | **53.98** | 46.71 | 69.75 | **84.70** | **53.91** | 54.37 |
|
70 |
+
| [OpenELM-270M-Instruct](https://huggingface.co/apple/OpenELM-270M-Instruct) | **30.55** | **46.68** | 48.56 | **52.07** | **70.78** | 84.40 | 52.72 | **55.11** |
|
71 |
+
| [OpenELM-450M](https://huggingface.co/apple/OpenELM-450M) | 27.56 | 48.06 | 55.78 | 53.97 | 72.31 | 87.20 | 58.01 | 57.56 |
|
72 |
+
| [OpenELM-450M-Instruct](https://huggingface.co/apple/OpenELM-450M-Instruct) | **30.38** | **50.00** | **60.37** | **59.34** | **72.63** | **88.00** | **58.96** | **59.95** |
|
73 |
+
| [OpenELM-1_1B](https://huggingface.co/apple/OpenELM-1_1B) | 32.34 | **55.43** | 63.58 | 64.81 | **75.57** | **90.60** | 61.72 | 63.44 |
|
74 |
+
| [OpenELM-1_1B-Instruct](https://huggingface.co/apple/OpenELM-1_1B-Instruct) | **37.97** | 52.23 | **70.00** | **71.20** | 75.03 | 89.30 | **62.75** | **65.50** |
|
75 |
+
| [OpenELM-3B](https://huggingface.co/apple/OpenELM-3B) | 35.58 | 59.89 | 67.40 | 72.44 | 78.24 | **92.70** | 65.51 | 67.39 |
|
76 |
+
| [OpenELM-3B-Instruct](https://huggingface.co/apple/OpenELM-3B-Instruct) | **39.42** | **61.74** | **68.17** | **76.36** | **79.00** | 92.50 | **66.85** | **69.15** |
|
77 |
+
|
78 |
+
### LLM360
|
79 |
+
|
80 |
+
| **Model Size** | **ARC-c** | **HellaSwag** | **MMLU** | **TruthfulQA** | **WinoGrande** | **Average** |
|
81 |
+
|-----------------------------------------------------------------------------|-----------|---------------|-----------|----------------|----------------|-------------|
|
82 |
+
| [OpenELM-270M](https://huggingface.co/apple/OpenELM-270M) | 27.65 | 47.15 | 25.72 | **39.24** | **53.83** | 38.72 |
|
83 |
+
| [OpenELM-270M-Instruct](https://huggingface.co/apple/OpenELM-270M-Instruct) | **32.51** | **51.58** | **26.70** | 38.72 | 53.20 | **40.54** |
|
84 |
+
| [OpenELM-450M](https://huggingface.co/apple/OpenELM-450M) | 30.20 | 53.86 | **26.01** | 40.18 | 57.22 | 41.50 |
|
85 |
+
| [OpenELM-450M-Instruct](https://huggingface.co/apple/OpenELM-450M-Instruct) | **33.53** | **59.31** | 25.41 | **40.48** | **58.33** | **43.41** |
|
86 |
+
| [OpenELM-1_1B](https://huggingface.co/apple/OpenELM-1_1B) | 36.69 | 65.71 | **27.05** | 36.98 | 63.22 | 45.93 |
|
87 |
+
| [OpenELM-1_1B-Instruct](https://huggingface.co/apple/OpenELM-1_1B-Instruct) | **41.55** | **71.83** | 25.65 | **45.95** | **64.72** | **49.94** |
|
88 |
+
| [OpenELM-3B](https://huggingface.co/apple/OpenELM-3B) | 42.24 | 73.28 | **26.76** | 34.98 | 67.25 | 48.90 |
|
89 |
+
| [OpenELM-3B-Instruct](https://huggingface.co/apple/OpenELM-3B-Instruct) | **47.70** | **76.87** | 24.80 | **38.76** | **67.96** | **51.22** |
|
90 |
+
|
91 |
+
|
92 |
+
### OpenLLM Leaderboard
|
93 |
+
|
94 |
+
| **Model Size** | **ARC-c** | **CrowS-Pairs** | **HellaSwag** | **MMLU** | **PIQA** | **RACE** | **TruthfulQA** | **WinoGrande** | **Average** |
|
95 |
+
|-----------------------------------------------------------------------------|-----------|-----------------|---------------|-----------|-----------|-----------|----------------|----------------|-------------|
|
96 |
+
| [OpenELM-270M](https://huggingface.co/apple/OpenELM-270M) | 27.65 | **66.79** | 47.15 | 25.72 | 69.75 | 30.91 | **39.24** | **53.83** | 45.13 |
|
97 |
+
| [OpenELM-270M-Instruct](https://huggingface.co/apple/OpenELM-270M-Instruct) | **32.51** | 66.01 | **51.58** | **26.70** | **70.78** | 33.78 | 38.72 | 53.20 | **46.66** |
|
98 |
+
| [OpenELM-450M](https://huggingface.co/apple/OpenELM-450M) | 30.20 | **68.63** | 53.86 | **26.01** | 72.31 | 33.11 | 40.18 | 57.22 | 47.69 |
|
99 |
+
| [OpenELM-450M-Instruct](https://huggingface.co/apple/OpenELM-450M-Instruct) | **33.53** | 67.44 | **59.31** | 25.41 | **72.63** | **36.84** | **40.48** | **58.33** | **49.25** |
|
100 |
+
| [OpenELM-1_1B](https://huggingface.co/apple/OpenELM-1_1B) | 36.69 | **71.74** | 65.71 | **27.05** | **75.57** | 36.46 | 36.98 | 63.22 | 51.68 |
|
101 |
+
| [OpenELM-1_1B-Instruct](https://huggingface.co/apple/OpenELM-1_1B-Instruct) | **41.55** | 71.02 | **71.83** | 25.65 | 75.03 | **39.43** | **45.95** | **64.72** | **54.40** |
|
102 |
+
| [OpenELM-3B](https://huggingface.co/apple/OpenELM-3B) | 42.24 | **73.29** | 73.28 | **26.76** | 78.24 | **38.76** | 34.98 | 67.25 | 54.35 |
|
103 |
+
| [OpenELM-3B-Instruct](https://huggingface.co/apple/OpenELM-3B-Instruct) | **47.70** | 72.33 | **76.87** | 24.80 | **79.00** | 38.47 | **38.76** | **67.96** | **55.73** |
|
104 |
+
|
105 |
+
See the technical report for more results and comparison.
|
106 |
+
|
107 |
+
## Evaluation
|
108 |
+
|
109 |
+
### Setup
|
110 |
+
|
111 |
+
Install the following dependencies:
|
112 |
+
|
113 |
+
```bash
|
114 |
+
|
115 |
+
# install public lm-eval-harness
|
116 |
+
|
117 |
+
harness_repo="public-lm-eval-harness"
|
118 |
+
git clone https://github.com/EleutherAI/lm-evaluation-harness ${harness_repo}
|
119 |
+
cd ${harness_repo}
|
120 |
+
# use main branch on 03-15-2024, SHA is dc90fec
|
121 |
+
git checkout dc90fec
|
122 |
+
pip install -e .
|
123 |
+
cd ..
|
124 |
+
|
125 |
+
# 66d6242 is the main branch on 2024-04-01
|
126 |
+
pip install datasets@git+https://github.com/huggingface/datasets.git@66d6242
|
127 |
+
pip install tokenizers>=0.15.2 transformers>=4.38.2 sentencepiece>=0.2.0
|
128 |
+
|
129 |
+
```
|
130 |
+
|
131 |
+
### Evaluate OpenELM
|
132 |
+
|
133 |
+
```bash
|
134 |
+
|
135 |
+
# OpenELM-450M-Instruct
|
136 |
+
hf_model=apple/OpenELM-450M-Instruct
|
137 |
+
|
138 |
+
# this flag is needed because lm-eval-harness set add_bos_token to False by default, but OpenELM uses LLaMA tokenizer which requires add_bos_token to be True
|
139 |
+
tokenizer=meta-llama/Llama-2-7b-hf
|
140 |
+
add_bos_token=True
|
141 |
+
batch_size=1
|
142 |
+
|
143 |
+
mkdir lm_eval_output
|
144 |
+
|
145 |
+
shot=0
|
146 |
+
task=arc_challenge,arc_easy,boolq,hellaswag,piqa,race,winogrande,sciq,truthfulqa_mc2
|
147 |
+
lm_eval --model hf \
|
148 |
+
--model_args pretrained=${hf_model},trust_remote_code=True,add_bos_token=${add_bos_token},tokenizer=${tokenizer} \
|
149 |
+
--tasks ${task} \
|
150 |
+
--device cuda:0 \
|
151 |
+
--num_fewshot ${shot} \
|
152 |
+
--output_path ./lm_eval_output/${hf_model//\//_}_${task//,/_}-${shot}shot \
|
153 |
+
--batch_size ${batch_size} 2>&1 | tee ./lm_eval_output/eval-${hf_model//\//_}_${task//,/_}-${shot}shot.log
|
154 |
+
|
155 |
+
shot=5
|
156 |
+
task=mmlu,winogrande
|
157 |
+
lm_eval --model hf \
|
158 |
+
--model_args pretrained=${hf_model},trust_remote_code=True,add_bos_token=${add_bos_token},tokenizer=${tokenizer} \
|
159 |
+
--tasks ${task} \
|
160 |
+
--device cuda:0 \
|
161 |
+
--num_fewshot ${shot} \
|
162 |
+
--output_path ./lm_eval_output/${hf_model//\//_}_${task//,/_}-${shot}shot \
|
163 |
+
--batch_size ${batch_size} 2>&1 | tee ./lm_eval_output/eval-${hf_model//\//_}_${task//,/_}-${shot}shot.log
|
164 |
+
|
165 |
+
shot=25
|
166 |
+
task=arc_challenge,crows_pairs_english
|
167 |
+
lm_eval --model hf \
|
168 |
+
--model_args pretrained=${hf_model},trust_remote_code=True,add_bos_token=${add_bos_token},tokenizer=${tokenizer} \
|
169 |
+
--tasks ${task} \
|
170 |
+
--device cuda:0 \
|
171 |
+
--num_fewshot ${shot} \
|
172 |
+
--output_path ./lm_eval_output/${hf_model//\//_}_${task//,/_}-${shot}shot \
|
173 |
+
--batch_size ${batch_size} 2>&1 | tee ./lm_eval_output/eval-${hf_model//\//_}_${task//,/_}-${shot}shot.log
|
174 |
+
|
175 |
+
shot=10
|
176 |
+
task=hellaswag
|
177 |
+
lm_eval --model hf \
|
178 |
+
--model_args pretrained=${hf_model},trust_remote_code=True,add_bos_token=${add_bos_token},tokenizer=${tokenizer} \
|
179 |
+
--tasks ${task} \
|
180 |
+
--device cuda:0 \
|
181 |
+
--num_fewshot ${shot} \
|
182 |
+
--output_path ./lm_eval_output/${hf_model//\//_}_${task//,/_}-${shot}shot \
|
183 |
+
--batch_size ${batch_size} 2>&1 | tee ./lm_eval_output/eval-${hf_model//\//_}_${task//,/_}-${shot}shot.log
|
184 |
+
|
185 |
+
```
|
186 |
+
|
187 |
+
|
188 |
+
## Bias, Risks, and Limitations
|
189 |
+
|
190 |
+
The release of OpenELM models aims to empower and enrich the open research community by providing access to state-of-the-art language models. Trained on publicly available datasets, these models are made available without any safety guarantees. Consequently, there exists the possibility of these models producing outputs that are inaccurate, harmful, biased, or objectionable in response to user prompts. Thus, it is imperative for users and developers to undertake thorough safety testing and implement appropriate filtering mechanisms tailored to their specific requirements.
|
191 |
+
|
192 |
+
## Citation
|
193 |
+
|
194 |
+
If you find our work useful, please cite:
|
195 |
+
|
196 |
+
```BibTex
|
197 |
+
@article{mehtaOpenELMEfficientLanguage2024,
|
198 |
+
title = {{OpenELM}: {An} {Efficient} {Language} {Model} {Family} with {Open} {Training} and {Inference} {Framework}},
|
199 |
+
shorttitle = {{OpenELM}},
|
200 |
+
url = {https://arxiv.org/abs/2404.14619v1},
|
201 |
+
language = {en},
|
202 |
+
urldate = {2024-04-24},
|
203 |
+
journal = {arXiv.org},
|
204 |
+
author = {Mehta, Sachin and Sekhavat, Mohammad Hossein and Cao, Qingqing and Horton, Maxwell and Jin, Yanzi and Sun, Chenfan and Mirzadeh, Iman and Najibi, Mahyar and Belenko, Dmitry and Zatloukal, Peter and Rastegari, Mohammad},
|
205 |
+
month = apr,
|
206 |
+
year = {2024},
|
207 |
+
}
|
208 |
+
|
209 |
+
@inproceedings{mehta2022cvnets,
|
210 |
+
author = {Mehta, Sachin and Abdolhosseini, Farzad and Rastegari, Mohammad},
|
211 |
+
title = {CVNets: High Performance Library for Computer Vision},
|
212 |
+
year = {2022},
|
213 |
+
booktitle = {Proceedings of the 30th ACM International Conference on Multimedia},
|
214 |
+
series = {MM '22}
|
215 |
+
}
|
216 |
+
```
|