Yukang commited on
Commit
5f03b06
1 Parent(s): 91fa326

Create README.md

Browse files
Files changed (1) hide show
  1. README.md +366 -0
README.md ADDED
@@ -0,0 +1,366 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # LongLoRA and LongAlpaca for Long-context LLMs
2
+
3
+
4
+ [![Huggingface Models](https://img.shields.io/badge/Models-Huggingface%20Models-bron)](https://huggingface.co/Yukang)
5
+ [![Github](https://img.shields.io/badge/Github-Repo-cyan)](https://github.com/dvlab-research/LongLoRA)
6
+ [![Data](https://img.shields.io/badge/Data-LongAlpaca%2012k-light)](https://huggingface.co/datasets/Yukang/LongAlpaca-12k)
7
+ [![Paper](https://img.shields.io/badge/Paper-Arvix-blue)](https://arxiv.org/abs/2309.12307)
8
+
9
+ [![Code License](https://img.shields.io/badge/Code%20License-Apache_2.0-yellow.svg)](https://github.com/dvlab-research/LongLoRA/blob/main/LICENSE)
10
+ [![Data License](https://img.shields.io/badge/Data%20License-CC%20By%20NC%204.0-orange.svg)](https://github.com/dvlab-research/LongLoRA/blob/main/DATA_LICENSE)
11
+ [![Weight License](https://img.shields.io/badge/Weight%20License-CC%20By%20NC%204.0-red)](https://github.com/dvlab-research/LongLoRA/blob/main/WEIGHT_LICENSE)
12
+
13
+ For detailed usage and codes, please visit the [Github project](https://github.com/dvlab-research/LongLoRA).
14
+ ## TABLE OF CONTENTS
15
+ 1. [News](#news)
16
+ 2. [Examples](#examples)
17
+ 3. [Highlights](#highlights)
18
+ 4. [How to contribute](#how-to-contribute)
19
+ 5. [Requirements](#usage-requirements)
20
+ 6. [Installation and quick guide](#installation-and-quick-guide)
21
+ 7. [LongAlpaca Data](#longalpaca-data)
22
+ 8. [Models](#models)
23
+ 9. [Training](#training)
24
+ 10. [Evaluation](#evaluation)
25
+ 11. [Demo](#demo)
26
+ 12. [Data Generation via Pdf2Text](#data-generation-via-pdf2text)
27
+ 13. [Citation](#citation)
28
+ 14. [Acknowledgement](#acknowledgement)
29
+ 15. [License](#license)
30
+
31
+ ## News
32
+ - [x] [2023.10.8] **We release the long instruction-following dataset**, [LongAlpaca-12k](https://huggingface.co/datasets/Yukang/LongAlpaca-12k) and **the corresponding models**, [LongAlpaca-7B](https://huggingface.co/Yukang/LongAlpaca-7B), [LongAlpaca-13B](https://huggingface.co/Yukang/LongAlpaca-13B), and [LongAlpaca-70B](https://huggingface.co/Yukang/LongAlpaca-70B).
33
+ - (*The previous sft models*, [Llama-2-13b-chat-longlora-32k-sft](https://huggingface.co/Yukang/Llama-2-13b-chat-longlora-32k-sft) and [Llama-2-70b-chat-longlora-32k-sft](https://huggingface.co/Yukang/Llama-2-70b-chat-longlora-32k-sft), *have been depreciated*.)
34
+ - [x] [2023.10.3] We add support GPTNeoX models. Please refer to this [PR](https://github.com/dvlab-research/LongLoRA/pull/32) for usage. Thanks for @naubull2 for this contribution.
35
+ - [x] [2023.9.22] We release all our fine-tuned [models](https://huggingface.co/Yukang), including **70B-32k models**, [LLaMA2-LongLoRA-70B-32k](https://huggingface.co/Yukang/Llama-2-70b-longlora-32k), [LLaMA2-LongLoRA-7B-100k](https://huggingface.co/Yukang/Llama-2-7b-longlora-100k-ft). Welcome to check them out!
36
+ - [x] [2023.9.22] We release [Paper](http://arxiv.org/abs/2309.12307) and this GitHub repo, including training and evaluation code.
37
+
38
+ **LongLoRA: Efficient Fine-tuning of Long-Context Large Language Models [[Paper](http://arxiv.org/abs/2309.12307)]** <br />
39
+ [Yukang Chen](https://scholar.google.com/citations?user=6p0ygKUAAAAJ&hl=en),
40
+ [Shengju Qian](https://scholar.google.com/citations?user=QNnWmasAAAAJ),
41
+ [Haotian Tang](https://scholar.google.com/citations?user=WxL13BAAAAAJ&hl),
42
+ [Xin Lai](https://scholar.google.com/citations?user=tqNDPA4AAAAJ&hl=zh-CN),
43
+ [Zhijian Liu](https://scholar.google.com/citations?user=3coYSTUAAAAJ&hl=en),
44
+ [Song Han](https://scholar.google.com/citations?user=E0iCaa4AAAAJ&hl=zh-CN),
45
+ [Jiaya Jia](https://scholar.google.com/citations?user=XPAkzTEAAAAJ&hl=en)<br />
46
+
47
+ ## Highlights
48
+ 1. In LongLoRA approach, The proposed shifted short attention is easy to implement, compatible with Flash-Attention, and is not required during inference.
49
+ 2. We released all our models, including models from 7B to 70B, context length from 8k to 100k, including [LLaMA2-LongLoRA-7B-100k](https://huggingface.co/Yukang/Llama-2-7b-longlora-100k-ft), [LLaMA2-LongLoRA-13B-64k](https://huggingface.co/Yukang/Llama-2-13b-longlora-64k), and [LLaMA2-LongLoRA-70B-32k](https://huggingface.co/Yukang/Llama-2-70b-longlora-32k).
50
+ 3. We built up a long-context instruction-following dataset, [LongAlpaca-12k](#longalpaca-data). We released the corresponding [LongAlpaca-7B](https://huggingface.co/Yukang/LongAlpaca-7B), [LongAlpaca-13B](https://huggingface.co/Yukang/LongAlpaca-13B) and [LongAlpaca-70B](https://huggingface.co/Yukang/LongAlpaca-70B) models. To our best knowledge, this is the first open-sourced long-context 70B model.
51
+
52
+ ## How to Contribute
53
+ - Make sure to have git installed.
54
+ - Create your own [fork](https://github.com/dvlab-research/LongLoRA/fork) of the project.
55
+ - Clone the repository on your local machine, using git clone and pasting the url of this project.
56
+ - Read both the `Requirements` and `Installation and Quick Guide` sections below.
57
+ - Commit and push your changes.
58
+ - Make a pull request when finished modifying the project.
59
+
60
+
61
+ ## Usage Requirements
62
+ To download and use the [pre-trained weights](#pre-trained-weights) you will need:
63
+ 1. Hugging Face (HF) account with valid email. Note, the email used for HF must alse be used for the license agreement.
64
+ 2. Accept the Meta [license and acceptable use policy](https://ai.meta.com/resources/models-and-libraries/llama-downloads/)
65
+
66
+
67
+ ## Installation and Quick Guide
68
+ To install and run the application:
69
+ 1. [Fork this repo](https://github.com/dvlab-research/LongLoRA/fork) on github
70
+ 2. Clone the repository on your local machine, using git clone and pasting the url of this project.
71
+ 3. Run the following code:
72
+ ```
73
+ pip install -r requirements.txt
74
+ pip install flash-attn --no-build-isolation
75
+ ```
76
+ 4. Use either a [Released model](#released-models) or [Fine tune](#fine-tuning) a model to fit your preferences.
77
+ 5. Test your model by chat.
78
+ 6. Deploy your own demo.
79
+
80
+ ## LongAlpaca Data
81
+
82
+ LongAlpaca-12k contains 9k long QA data that we collected and 3k short QA sampled from the original [Alpaca data](https://github.com/tatsu-lab/stanford_alpaca/blob/main/alpaca_data.json). This is to avoid the case that the model might degrade at short instruction following. The data we collect contains various types and amounts as the following figure.
83
+
84
+ | Data | Short QA | Long QA | Total | Download |
85
+ |:---------------|----------|----------|----------|----------|
86
+ | LongAlpaca-12k | 3k | 9k | 12k | [Link](https://huggingface.co/datasets/Yukang/LongAlpaca-12k) |
87
+
88
+ Following the original Alpaca format, our Long QA data uses the following prompts for fine-tuning:
89
+ - `instruction`: `str`, describes the task the model should perform. For example, to answer a question after reading a book section or paper. We vary the contents and questions to make instructions diverse.
90
+ - `output`: `str`, the answer to the instruction.
91
+
92
+ We did not use the `input` format in the Alpaca format for simplicity.
93
+
94
+ ## Models
95
+
96
+ ### Models with supervised fine-tuning
97
+ | Model | Size | Context | Train | Link |
98
+ |:---------------|------|---------|---------|-----------------------------------------------------------------------------------------------------------------------|
99
+ | LongAlpaca-7B | 7B | 32768 | Full FT | [Model](https://huggingface.co/Yukang/LongAlpaca-7B) |
100
+ | LongAlpaca-13B | 13B | 32768 | Full FT | [Model](https://huggingface.co/Yukang/LongAlpaca-13B) |
101
+ | LongAlpaca-70B | 70B | 32768 | LoRA+ | [Model](https://huggingface.co/Yukang/LongAlpaca-70B) [(LoRA-weight)](https://huggingface.co/Yukang/LongAlpaca-70B-lora) |
102
+
103
+
104
+ ### Models with context extension via fully fine-tuning
105
+ | Model | Size | Context | Train | Link |
106
+ |:----------------------------|------|---------|-------|-------------------------------------------------------------------|
107
+ | Llama-2-7b-longlora-8k-ft | 7B | 8192 | Full FT | [Model](https://huggingface.co/Yukang/Llama-2-7b-longlora-8k-ft) |
108
+ | Llama-2-7b-longlora-16k-ft | 7B | 16384 | Full FT | [Model](https://huggingface.co/Yukang/Llama-2-7b-longlora-16k-ft) |
109
+ | Llama-2-7b-longlora-32k-ft | 7B | 32768 | Full FT | [Model](https://huggingface.co/Yukang/Llama-2-7b-longlora-32k-ft) |
110
+ | Llama-2-7b-longlora-100k-ft | 7B | 100000 | Full FT | [Model](https://huggingface.co/Yukang/Llama-2-7b-longlora-100k-ft) |
111
+ | Llama-2-13b-longlora-8k-ft | 13B | 8192 | Full FT | [Model](https://huggingface.co/Yukang/Llama-2-13b-longlora-8k-ft) |
112
+ | Llama-2-13b-longlora-16k-ft | 13B | 16384 | Full FT | [Model](https://huggingface.co/Yukang/Llama-2-13b-longlora-16k-ft) |
113
+ | Llama-2-13b-longlora-32k-ft | 13B | 32768 | Full FT | [Model](https://huggingface.co/Yukang/Llama-2-13b-longlora-32k-ft) |
114
+
115
+ ### Models with context extension via improved LoRA fine-tuning
116
+ | Model | Size | Context | Train | Link |
117
+ |:----------------------------|------|---------|-------|---------------------------------------------------------------------|
118
+ | Llama-2-7b-longlora-8k | 7B | 8192 | LoRA+ | [LoRA-weight](https://huggingface.co/Yukang/Llama-2-7b-longlora-8k) |
119
+ | Llama-2-7b-longlora-16k | 7B | 16384 | LoRA+ | [LoRA-weight](https://huggingface.co/Yukang/Llama-2-7b-longlora-16k) |
120
+ | Llama-2-7b-longlora-32k | 7B | 32768 | LoRA+ | [LoRA-weight](https://huggingface.co/Yukang/Llama-2-7b-longlora-32k) |
121
+ | Llama-2-13b-longlora-8k | 13B | 8192 | LoRA+ | [LoRA-weight](https://huggingface.co/Yukang/Llama-2-13b-longlora-8k) |
122
+ | Llama-2-13b-longlora-16k | 13B | 16384 | LoRA+ | [LoRA-weight](https://huggingface.co/Yukang/Llama-2-13b-longlora-16k) |
123
+ | Llama-2-13b-longlora-32k | 13B | 32768 | LoRA+ | [LoRA-weight](https://huggingface.co/Yukang/Llama-2-13b-longlora-32k) |
124
+ | Llama-2-13b-longlora-64k | 13B | 65536 | LoRA+ | [LoRA-weight](https://huggingface.co/Yukang/Llama-2-13b-longlora-64k) |
125
+ | Llama-2-70b-longlora-32k | 70B | 32768 | LoRA+ | [LoRA-weight](https://huggingface.co/Yukang/Llama-2-70b-longlora-32k) |
126
+ | Llama-2-70b-chat-longlora-32k | 70B | 32768 | LoRA+ | [LoRA-weight](https://huggingface.co/Yukang/Llama-2-70b-chat-longlora-32k) |
127
+
128
+ ## Training
129
+ ### Pre-trained weights
130
+ We use LLaMA2 models as the pre-trained weights and fine-tune them to long context window sizes. Download based on your choices.
131
+
132
+ | Pre-trained weights |
133
+ |:-------------------------------------------------------------------------------------|
134
+ | [Llama-2-7b-hf](https://huggingface.co/meta-llama/Llama-2-7b-hf) |
135
+ |[Llama-2-13b-hf](https://huggingface.co/meta-llama/Llama-2-13b-hf) |
136
+ | [Llama-2-70b-hf](https://huggingface.co/meta-llama/Llama-2-70b-hf) |
137
+ | [Llama-2-7b-chat-hf](https://huggingface.co/meta-llama/Llama-2-7b-chat-hf) |
138
+ | [Llama-2-13b-chat-hf](https://huggingface.co/meta-llama/Llama-2-13b-chat-hf) |
139
+ | [Llama-2-70b-chat-hf](https://huggingface.co/meta-llama/Llama-2-70b-chat-hf) |
140
+
141
+ This project also supports GPTNeoX models as the base model architecture. Some candidate pre-trained weights may include [GPT-NeoX-20B](https://huggingface.co/EleutherAI/gpt-neox-20b), [Polyglot-ko-12.8B](https://huggingface.co/EleutherAI/polyglot-ko-12.8b) and other variants.
142
+
143
+ ### Fine-tuning
144
+ ```
145
+ torchrun --nproc_per_node=8 fine-tune.py \
146
+ --model_name_or_path path_to/Llama-2-7b-hf \
147
+ --bf16 True \
148
+ --output_dir path_to_saving_checkpoints \
149
+ --cache_dir path_to_cache \
150
+ --model_max_length 8192 \
151
+ --use_flash_attn True \
152
+ --low_rank_training False \
153
+ --num_train_epochs 1 \
154
+ --per_device_train_batch_size 1 \
155
+ --per_device_eval_batch_size 2 \
156
+ --gradient_accumulation_steps 8 \
157
+ --evaluation_strategy "no" \
158
+ --save_strategy "steps" \
159
+ --save_steps 1000 \
160
+ --save_total_limit 2 \
161
+ --learning_rate 2e-5 \
162
+ --weight_decay 0.0 \
163
+ --warmup_steps 20 \
164
+ --lr_scheduler_type "constant_with_warmup" \
165
+ --logging_steps 1 \
166
+ --deepspeed "ds_configs/stage2.json" \
167
+ --tf32 True \
168
+ --max_steps 1000
169
+ ```
170
+
171
+ - Please remember to change `path_to/Llama-2-7b-hf`, `path_to_saving_checkpoints`, `path_to_cache` to your own directory.
172
+ - Note that you can change `model_max_length` to other values.
173
+ - You could change `ds_configs/stage2.json` to `ds_configs/stage3.json` if you want.
174
+ - Please set `use_flash_attn` as `False` if you use V100 machines or do not install flash attention.
175
+ - You can set `low_rank_training` as `False` if you want to use fully fine-tuning. It will cost more GPU memory and slower, but the performance will be a bit better.
176
+ - When training is finished, to get the full model weight:
177
+ ```
178
+ cd path_to_saving_checkpoints && python zero_to_fp32.py . pytorch_model.bin
179
+ ```
180
+
181
+ ### Supervised Fine-tuning
182
+ ```
183
+ torchrun --nproc_per_node=8 supervised-fine-tune.py \
184
+ --model_name_or_path path_to_Llama2_chat_models \
185
+ --bf16 True \
186
+ --output_dir path_to_saving_checkpoints \
187
+ --model_max_length 32768 \
188
+ --use_flash_attn True \
189
+ --data_path LongAlpaca-12k.json \
190
+ --low_rank_training True \
191
+ --num_train_epochs 3 \
192
+ --per_device_train_batch_size 1 \
193
+ --per_device_eval_batch_size 2 \
194
+ --gradient_accumulation_steps 1 \
195
+ --evaluation_strategy "no" \
196
+ --save_strategy "steps" \
197
+ --save_steps 1000 \
198
+ --save_total_limit 2 \
199
+ --learning_rate 2e-5 \
200
+ --weight_decay 0.0 \
201
+ --warmup_steps 20 \
202
+ --lr_scheduler_type "constant_with_warmup" \
203
+ --logging_steps 1 \
204
+ --deepspeed "ds_configs/stage2.json" \
205
+ --tf32 True
206
+ ```
207
+ - There is no need to make supervised fine-tuning upon the fine-tuned context extended models. It is all right to directly use base model as Llama2-chat models, as the amount of long instruction following data is enough for SFT.
208
+ - Our long instruction following data can be found in [LongAlpaca-12k.json](https://huggingface.co/datasets/Yukang/LongAlpaca-12k).
209
+
210
+
211
+ ### Get trainable weights in low-rank training
212
+ In low-rank training, we set embedding and normalization layers as trainable. Please use the following line to extract the trainable weights `trainable_params.bin` from `pytorch_model.bin`
213
+ ```
214
+ python3 get_trainable_weights.py --checkpoint_path path_to_saving_checkpoints --trainable_params "embed,norm"
215
+ ```
216
+
217
+ ### Merge LoRA Weight
218
+ Merge the LoRA weights of `pytorch_model.bin` and trainable parameters `trainable_params.bin`, save the resulting model into your desired path in the Hugging Face format:
219
+ ```
220
+ python3 merge_lora_weights_and_save_hf_model.py \
221
+ --base_model path_to/Llama-2-7b-hf \
222
+ --peft_model path_to_saving_checkpoints \
223
+ --context_size 8192 \
224
+ --save_path path_to_saving_merged_model
225
+ ```
226
+ For example,
227
+ ```
228
+ python3 merge_lora_weights_and_save_hf_model.py \
229
+ --base_model /dataset/pretrained-models/Llama-2-7b-hf \
230
+ --peft_model /dataset/yukangchen/hf_models/lora-models/Llama-2-7b-longlora-8k \
231
+ --context_size 8192 \
232
+ --save_path /dataset/yukangchen/models/Llama-2-7b-longlora-8k-merged
233
+ ```
234
+
235
+
236
+ ## Evaluation
237
+ ### Perplexity Validation
238
+ To evaluate a model that is trained in the low-rank setting, please set both `base_model` and `peft_model`. `base_model` is the pre-trained weight. `peft_model` is the path to the saved checkpoint, which should contain `trainable_params.bin`, `adapter_model.bin` and `adapter_config.json`. For example,
239
+ ```
240
+ python3 eval.py --seq_len 8192 --context_size 8192 --batch_size 1 --base_model path_to/Llama-2-7b-hf --peft_model path_to_saving_checkpoints --data_path pg19/test.bin
241
+ ```
242
+
243
+ To evaluate a model that is fully fine-tuned, you only need to set `base_model` as the path to the saved checkpoint, which should contain `pytorch_model.bin` and `config.json`. `peft_model` should be ignored.
244
+ ```
245
+ python3 eval.py --seq_len 8192 --context_size 8192 --batch_size 1 --base_model path_to_saving_checkpoints --data_path pg19/test.bin
246
+ ```
247
+
248
+ - Note that `--seq_len` is to set the sequence length for evaluation. `--context_size` is to set the context length of the model during fine-tuning. `--seq_len` should not be larger than `--context_size`.
249
+
250
+ - We have already tokenized the validation and test splits of PG19 and proof-pile dataset into `pg19/validation.bin`, `pg19/test.bin`, and `proof-pile/test_sampled_data.bin`, with the tokenizer of LLaMA. `proof-pile/test_sampled_data.bin` contains 128 documents that are randomly sampled from the total proof-pile test split. For each document, it has at least 32768 tokens. We also release the sampled ids in [proof-pile/test_sampled_ids.bin](https://drive.google.com/file/d/1cnzWODLRQYAd7HeugzLCIhaqzaLZv7J5/view?usp=share_link). You can download them from the links below.
251
+
252
+ | Dataset | Split | Link |
253
+ |:-----------|------------|--------------------------------------------------------------------------------------------------------------|
254
+ | PG19 | validation | [pg19/validation.bin](https://drive.google.com/file/d/1rbJvb0qRIf2mQoN2ON7S93TbTzMnlrN6/view?usp=share_link) |
255
+ | PG19 | test | [pg19/test.bin](https://drive.google.com/file/d/1QANDMdctpacPAYgS04adDXqByGEq-Ret/view?usp=share_link) |
256
+ | Proof-pile | test | [proof-pile/test_sampled_data.bin](https://drive.google.com/file/d/1bUI5lPDvrqzY_XXJJ2sSuvZx0Y9AZClE/view?usp=share_link) |
257
+
258
+
259
+ ### Passkey Retrieval
260
+ We provide a manner to test the passkey retrieval accuracy. For example,
261
+ ```
262
+ python3 passkey_retrivial.py \
263
+ --context_size 32768 \
264
+ --base_model path_to/Llama-2-7b-longlora-32k \
265
+ --max_tokens 32768 \
266
+ --interval 1000
267
+ ```
268
+ - Note that the `context_size` is the context length during fine-tuning.
269
+ - `max_tokens` is maximum length for the document in passkey retrieval evaluation.
270
+ - `interval` is the interval during the document length increasing. It is a rough number because the document increases by sentences.
271
+
272
+ ## Demo
273
+ ### Local Inference
274
+ To chat with [Llama-2-13b-chat-longlora-32k-sft](https://huggingface.co/Yukang/Llama-2-13b-chat-longlora-32k-sft) or [Llama-2-70b-chat-longlora-32k-sft](https://huggingface.co/Yukang/Llama-2-70b-chat-longlora-32k-sft), you need to run `merge_lora_weights_and_save_hf_model.py` first, and then:
275
+ ```
276
+ python3 inference.py \
277
+ --base_model path_to_model \
278
+ --question $question \
279
+ --context_size $context_length \
280
+ --max_gen_len $max_gen_len \
281
+ --flash_attn True \
282
+ --material $material_content \
283
+ --material_type $material_type \
284
+ --material_title $material_title
285
+ ```
286
+ To ask a question related to a book:
287
+ ```
288
+ python3 inference.py \
289
+ --base_model /data/models/Llama-2-13b-chat-longlora-32k-sft \
290
+ --question "Why doesn't Professor Snape seem to like Harry?" \
291
+ --context_size 32768 \
292
+ --max_gen_len 512 \
293
+ --flash_attn True \
294
+ --material "materials/Harry Potter and the Philosophers Stone_section2.txt" \
295
+ --material_type "book" \
296
+ --material_title "Harry Potter and the Philosophers Stone"
297
+ ```
298
+ Note that you can ignore `material_type` or `material_title`.
299
+
300
+ To ask a question related to a paper:
301
+ ```
302
+ python3 inference.py \
303
+ --base_model /data/models/Llama-2-13b-chat-longlora-32k-sft \
304
+ --question "What are the main contributions and novelties of this work?" \
305
+ --context_size 32768 \
306
+ --max_gen_len 512 \
307
+ --flash_attn True \
308
+ --material "materials/paper1.txt" \
309
+ --material_type "paper"
310
+ ```
311
+
312
+ ### Online Demo
313
+ To deploy your own demo run
314
+ ```
315
+ python3 demo.py \
316
+ --base_model path_to_model \
317
+ --context_size $context_size \
318
+ --max_gen_len $max_gen_len \
319
+ --flash_attn True
320
+ ```
321
+ Example
322
+ ```
323
+ python3 demo.py \
324
+ --base_model /data/models/Llama-2-13b-chat-longlora-32k-sft \
325
+ --context_size 32768 \
326
+ --max_gen_len 512 \
327
+ --flash_attn True
328
+ ```
329
+ - Note that `flash_attn=True` will make the generation slow but save much GPU memory.
330
+
331
+ ## Data Generation via Pdf2text
332
+ During our dataset collection, we convert paper and books from pdf to text. The conversion quality has a large influence on the final model quality. We think that this step is non-trivial. We release the tool for the pdf2txt conversion, in the folder `pdf2txt`. It is built upon `pdf2image`, `easyocr`, `ditod` and `detectron2`. Please refer to the [README.md](pdf2txt/README.md) in `pdf2txt` for more details.
333
+
334
+ ## Citation
335
+ If you find this project useful in your research, please consider citing:
336
+
337
+ ```
338
+ @article{longlora,
339
+ title={LongLoRA: Efficient Fine-tuning of Long-Context Large Language Models},
340
+ author={Yukang Chen and Shengju Qian and Haotian Tang and Xin Lai and Zhijian Liu and Song Han and Jiaya Jia},
341
+ journal={arXiv:2309.12307},
342
+ year={2023}
343
+ }
344
+ ```
345
+
346
+
347
+ ```
348
+ @misc{long-alpaca,
349
+ author = {Yukang Chen and Shaozuo Yu and Shengju Qian and Haotian Tang and Xin Lai and Zhijian Liu and Song Han and Jiaya Jia},
350
+ title = {Long Alpaca: Long-context Instruction-following models},
351
+ year = {2023},
352
+ publisher = {GitHub},
353
+ journal = {GitHub repository},
354
+ howpublished = {\url{https://github.com/dvlab-research/LongLoRA}},
355
+ }
356
+ ```
357
+ ## Acknowledgement
358
+ - This work is built upon the [LLaMA2](https://ai.meta.com/llama) as the pre-trained models.
359
+ - This work can also be built upon the [GPTNeoX-HF](https://huggingface.co/docs/transformers/model_doc/gpt_neox) which is based upon [EleutherAI/GPTNeoX](https://github.com/EleutherAI/gpt-neox) as the pre-trained model architecture.
360
+ - This work is based on [DeepSpeed](https://github.com/microsoft/DeepSpeed), [peft](https://github.com/huggingface/peft), and [Flash-Attention2](https://github.com/Dao-AILab/flash-attention) for acceleration.
361
+ - Some evaluation code is modified upon [Landmark Attention](https://github.com/epfml/landmark-attention).
362
+ - We use [LongChat](https://github.com/DachengLi1/LongChat) for the retrieval evaluation.
363
+
364
+ ## License
365
+ - LongLoRA is licensed under the Apache License 2.0. This means that it requires the preservation of copyright and license notices.
366
+ - Data and weights are under CC-BY-NC 4.0 License. They are licensed for research use only, and allowed only non-commercial. Models trained using the dataset should not be used outside of research purposes.