raffr commited on
Commit
7f51a94
0 Parent(s):

Duplicate from localmodels/LLM

Browse files
.gitattributes ADDED
@@ -0,0 +1,34 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ *.7z filter=lfs diff=lfs merge=lfs -text
2
+ *.arrow filter=lfs diff=lfs merge=lfs -text
3
+ *.bin filter=lfs diff=lfs merge=lfs -text
4
+ *.bz2 filter=lfs diff=lfs merge=lfs -text
5
+ *.ckpt filter=lfs diff=lfs merge=lfs -text
6
+ *.ftz filter=lfs diff=lfs merge=lfs -text
7
+ *.gz filter=lfs diff=lfs merge=lfs -text
8
+ *.h5 filter=lfs diff=lfs merge=lfs -text
9
+ *.joblib filter=lfs diff=lfs merge=lfs -text
10
+ *.lfs.* filter=lfs diff=lfs merge=lfs -text
11
+ *.mlmodel filter=lfs diff=lfs merge=lfs -text
12
+ *.model filter=lfs diff=lfs merge=lfs -text
13
+ *.msgpack filter=lfs diff=lfs merge=lfs -text
14
+ *.npy filter=lfs diff=lfs merge=lfs -text
15
+ *.npz filter=lfs diff=lfs merge=lfs -text
16
+ *.onnx filter=lfs diff=lfs merge=lfs -text
17
+ *.ot filter=lfs diff=lfs merge=lfs -text
18
+ *.parquet filter=lfs diff=lfs merge=lfs -text
19
+ *.pb filter=lfs diff=lfs merge=lfs -text
20
+ *.pickle filter=lfs diff=lfs merge=lfs -text
21
+ *.pkl filter=lfs diff=lfs merge=lfs -text
22
+ *.pt filter=lfs diff=lfs merge=lfs -text
23
+ *.pth filter=lfs diff=lfs merge=lfs -text
24
+ *.rar filter=lfs diff=lfs merge=lfs -text
25
+ *.safetensors filter=lfs diff=lfs merge=lfs -text
26
+ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
27
+ *.tar.* filter=lfs diff=lfs merge=lfs -text
28
+ *.tflite filter=lfs diff=lfs merge=lfs -text
29
+ *.tgz filter=lfs diff=lfs merge=lfs -text
30
+ *.wasm filter=lfs diff=lfs merge=lfs -text
31
+ *.xz filter=lfs diff=lfs merge=lfs -text
32
+ *.zip filter=lfs diff=lfs merge=lfs -text
33
+ *.zst filter=lfs diff=lfs merge=lfs -text
34
+ *tfevents* filter=lfs diff=lfs merge=lfs -text
README.md ADDED
@@ -0,0 +1,211 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ duplicated_from: localmodels/LLM
3
+ ---
4
+ # Guanaco 33B ggml
5
+
6
+ From: https://huggingface.co/timdettmers/guanaco-33b-merged
7
+
8
+ ---
9
+
10
+ ### Original llama.cpp quant methods: `q4_0, q4_1, q5_0, q5_1, q8_0`
11
+
12
+ Quantized using an older version of llama.cpp and compatible with llama.cpp from May 19, commit 2d5db48.
13
+
14
+ ### k-quant methods: `q2_K, q3_K_S, q3_K_M, q3_K_L, q4_K_S, q4_K_M, q5_K_S, q6_K`
15
+
16
+ Quantization methods compatible with latest llama.cpp from June 6, commit 2d43387.
17
+
18
+ ---
19
+
20
+ ## Files
21
+ | Name | Quant method | Bits | Size | Max RAM required, no GPU offloading | Use case |
22
+ | ---- | ---- | ---- | ---- | ---- | ----- |
23
+ | guanaco-33B.ggmlv3.q2_K.bin | q2_K | 2 | 13.60 GB | 16.10 GB | New k-quant method. Uses GGML_TYPE_Q4_K for the attention.vw and feed_forward.w2 tensors, GGML_TYPE_Q2_K for the other tensors. |
24
+ | guanaco-33B.ggmlv3.q3_K_L.bin | q3_K_L | 3 | 17.20 GB | 19.70 GB | New k-quant method. Uses GGML_TYPE_Q5_K for the attention.wv, attention.wo, and feed_forward.w2 tensors, else GGML_TYPE_Q3_K |
25
+ | guanaco-33B.ggmlv3.q3_K_M.bin | q3_K_M | 3 | 15.64 GB | 18.14 GB | New k-quant method. Uses GGML_TYPE_Q4_K for the attention.wv, attention.wo, and feed_forward.w2 tensors, else GGML_TYPE_Q3_K |
26
+ | guanaco-33B.ggmlv3.q3_K_S.bin | q3_K_S | 3 | 13.98 GB | 16.48 GB | New k-quant method. Uses GGML_TYPE_Q3_K for all tensors |
27
+ | guanaco-33B.ggmlv3.q4_0.bin | q4_0 | 4 | 18.30 GB | 20.80 GB | Original llama.cpp quant method, 4-bit. |
28
+ | guanaco-33B.ggmlv3.q4_1.bin | q4_1 | 4 | 20.33 GB | 22.83 GB | Original llama.cpp quant method, 4-bit. Higher accuracy than q4_0 but not as high as q5_0. However has quicker inference than q5 models. |
29
+ | guanaco-33B.ggmlv3.q4_K_M.bin | q4_K_M | 4 | 19.57 GB | 22.07 GB | New k-quant method. Uses GGML_TYPE_Q6_K for half of the attention.wv and feed_forward.w2 tensors, else GGML_TYPE_Q4_K |
30
+ | guanaco-33B.ggmlv3.q4_K_S.bin | q4_K_S | 4 | 18.30 GB | 20.80 GB | New k-quant method. Uses GGML_TYPE_Q4_K for all tensors |
31
+ | guanaco-33B.ggmlv3.q5_0.bin | q5_0 | 5 | 22.37 GB | 24.87 GB | Original llama.cpp quant method, 5-bit. Higher accuracy, higher resource usage and slower inference. |
32
+ | guanaco-33B.ggmlv3.q5_1.bin | q5_1 | 5 | 24.40 GB | 26.90 GB | Original llama.cpp quant method, 5-bit. Even higher accuracy, resource usage and slower inference. |
33
+ | guanaco-33B.ggmlv3.q5_K_M.bin | q5_K_M | 5 | 23.02 GB | 25.52 GB | New k-quant method. Uses GGML_TYPE_Q6_K for half of the attention.wv and feed_forward.w2 tensors, else GGML_TYPE_Q5_K |
34
+ | guanaco-33B.ggmlv3.q5_K_S.bin | q5_K_S | 5 | 22.37 GB | 24.87 GB | New k-quant method. Uses GGML_TYPE_Q5_K for all tensors |
35
+ | guanaco-33B.ggmlv3.q6_K.bin | q6_K | 6 | 26.69 GB | 29.19 GB | New k-quant method. Uses GGML_TYPE_Q8_K - 6-bit quantization - for all tensors |
36
+ | guanaco-33B.ggmlv3.q8_0.bin | q8_0 | 8 | 34.56 GB | 37.06 GB | Original llama.cpp quant method, 8-bit. Almost indistinguishable from float16. High resource use and slow. Not recommended for most users. |
37
+
38
+ ---
39
+
40
+ # Guanaco Models Based on LLaMA
41
+
42
+ | [Paper](https://arxiv.org/abs/2305.14314) | [Code](https://github.com/artidoro/qlora) | [Demo](https://huggingface.co/spaces/uwnlp/guanaco-playground-tgi) |
43
+
44
+ **The Guanaco models are open-source finetuned chatbots obtained through 4-bit QLoRA tuning of LLaMA base models on the OASST1 dataset. They are available in 7B, 13B, 33B, and 65B parameter sizes.**
45
+
46
+ ⚠️Guanaco is a model purely intended for research purposes and could produce problematic outputs.
47
+
48
+ ## Why use Guanaco?
49
+ - **Competitive with commercial chatbot systems on the Vicuna and OpenAssistant benchmarks** (ChatGPT and BARD) according to human and GPT-4 raters. We note that the relative performance on tasks not covered in these benchmarks could be very different. In addition, commercial systems evolve over time (we used outputs from the March 2023 version of the models).
50
+ - **Available open-source for research purposes**. Guanaco models allow *cheap* and *local* experimentation with high-quality chatbot systems.
51
+ - **Replicable and efficient training procedure** that can be extended to new use cases. Guanaco training scripts are available in the [QLoRA repo](https://github.com/artidoro/qlora).
52
+ - **Rigorous comparison to 16-bit methods** (both 16-bit full-finetuning and LoRA) in [our paper](https://arxiv.org/abs/2305.14314) demonstrates the effectiveness of 4-bit QLoRA finetuning.
53
+ - **Lightweight** checkpoints which only contain adapter weights.
54
+
55
+ ## License and Intended Use
56
+ Guanaco adapter weights are available under Apache 2 license. Note the use of the Guanaco adapter weights, requires access to the LLaMA model weighs.
57
+ Guanaco is based on LLaMA and therefore should be used according to the LLaMA license.
58
+
59
+ ## Usage
60
+ Here is an example of how you would load Guanaco 7B in 4-bits:
61
+ ```python
62
+ import torch
63
+ from peft import PeftModel
64
+ from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
65
+
66
+ model_name = "huggyllama/llama-7b"
67
+ adapters_name = 'timdettmers/guanaco-7b'
68
+
69
+ model = AutoModelForCausalLM.from_pretrained(
70
+ model_name,
71
+ load_in_4bit=True,
72
+ torch_dtype=torch.bfloat16,
73
+ device_map="auto",
74
+ max_memory= {i: '24000MB' for i in range(torch.cuda.device_count())},
75
+ quantization_config=BitsAndBytesConfig(
76
+ load_in_4bit=True,
77
+ bnb_4bit_compute_dtype=torch.bfloat16,
78
+ bnb_4bit_use_double_quant=True,
79
+ bnb_4bit_quant_type='nf4'
80
+ ),
81
+ )
82
+ model = PeftModel.from_pretrained(model, adapters_name)
83
+ tokenizer = AutoTokenizer.from_pretrained(model_name)
84
+
85
+ ```
86
+ Inference can then be performed as usual with HF models as follows:
87
+ ```python
88
+ prompt = "Introduce yourself"
89
+ formatted_prompt = (
90
+ f"A chat between a curious human and an artificial intelligence assistant."
91
+ f"The assistant gives helpful, detailed, and polite answers to the user's questions.\n"
92
+ f"### Human: {prompt} ### Assistant:"
93
+ )
94
+ inputs = tokenizer(formatted_prompt, return_tensors="pt").to("cuda:0")
95
+ outputs = model.generate(inputs=inputs.input_ids, max_new_tokens=20)
96
+ print(tokenizer.decode(outputs[0], skip_special_tokens=True))
97
+ ```
98
+ Expected output similar to the following:
99
+ ```
100
+ A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions.
101
+ ### Human: Introduce yourself ### Assistant: I am an artificial intelligence assistant. I am here to help you with any questions you may have.
102
+ ```
103
+
104
+
105
+ ## Current Inference Limitations
106
+ Currently, 4-bit inference is slow. We recommend loading in 16 bits if inference speed is a concern. We are actively working on releasing efficient 4-bit inference kernels.
107
+
108
+ Below is how you would load the model in 16 bits:
109
+ ```python
110
+ model_name = "huggyllama/llama-7b"
111
+ adapters_name = 'timdettmers/guanaco-7b'
112
+ model = AutoModelForCausalLM.from_pretrained(
113
+ model_name,
114
+ torch_dtype=torch.bfloat16,
115
+ device_map="auto",
116
+ max_memory= {i: '24000MB' for i in range(torch.cuda.device_count())},
117
+ )
118
+ model = PeftModel.from_pretrained(model, adapters_name)
119
+ tokenizer = AutoTokenizer.from_pretrained(model_name)
120
+
121
+ ```
122
+
123
+
124
+ ## Model Card
125
+ **Architecture**: The Guanaco models are LoRA adapters to be used on top of LLaMA models. They are added to all layers. For all model sizes, we use $r=64$.
126
+
127
+ **Base Model**: Guanaco uses LLaMA as base model with sizes 7B, 13B, 33B, 65B. LLaMA is a causal language model pretrained on a large corpus of text. See [LLaMA paper](https://arxiv.org/abs/2302.13971) for more details. Note that Guanaco can inherit biases and limitations of the base model.
128
+
129
+ **Finetuning Data**: Guanaco is finetuned on OASST1. The exact dataset is available at [timdettmers/openassistant-guanaco](https://huggingface.co/datasets/timdettmers/openassistant-guanaco).
130
+
131
+ **Languages**: The OASST1 dataset is multilingual (see [the paper](https://arxiv.org/abs/2304.07327) for details) and as such Guanaco responds to user queries in different languages. We note, however, that OASST1 is heavy in high-resource languages. In addition, human evaluation of Guanaco was only performed in English and based on qualitative analysis we observed degradation in performance in other languages.
132
+
133
+ Next, we describe Training and Evaluation details.
134
+
135
+ ### Training
136
+ Guanaco models are the result of 4-bit QLoRA supervised finetuning on the OASST1 dataset.
137
+
138
+ All models use NormalFloat4 datatype for the base model and LoRA adapters on all linear layers with BFloat16 as computation datatype. We set LoRA $r=64$, $\alpha=16$. We also use Adam beta2 of 0.999, max grad norm of 0.3 and LoRA dropout of 0.1 for models up to 13B and 0.05 for 33B and 65B models.
139
+ For the finetuning process, we use constant learning rate schedule and paged AdamW optimizer.
140
+
141
+ ### Training hyperparameters
142
+ Size| Dataset | Batch Size | Learning Rate | Max Steps | Sequence length
143
+ ---|---|---|---|---|---
144
+ 7B | OASST1 | 16 | 2e-4 | 1875 | 512
145
+ 13B | OASST1 | 16 | 2e-4 | 1875 | 512
146
+ 33B | OASST1 | 16 | 1e-4 | 1875 | 512
147
+ 65B | OASST1 | 16 | 1e-4 | 1875 | 512
148
+
149
+ ### Evaluation
150
+ We test generative language capabilities through both automated and human evaluations. This second set of evaluations relies on queries curated by humans and aims at measuring the quality of model responses. We use the Vicuna and OpenAssistant datasets with 80 and 953 prompts respectively.
151
+
152
+ In both human and automated evaluations, for each prompt, raters compare all pairs of responses across the models considered. For human raters we randomize the order of the systems, for GPT-4 we evaluate with both orders.
153
+
154
+
155
+ Benchmark | Vicuna | | Vicuna | | OpenAssistant | | -
156
+ -----------|----|-----|--------|---|---------------|---|---
157
+ Prompts | 80 | | 80 | | 953 | |
158
+ Judge | Human | | GPT-4 | | GPT-4 | |
159
+ Model | Elo | Rank | Elo | Rank | Elo | Rank | **Median Rank**
160
+ GPT-4 | 1176 | 1 | 1348 | 1 | 1294 | 1 | 1
161
+ Guanaco-65B | 1023 | 2 | 1022 | 2 | 1008 | 3 | 2
162
+ Guanaco-33B | 1009 | 4 | 992 | 3 | 1002 | 4 | 4
163
+ ChatGPT-3.5 Turbo | 916 | 7 | 966 | 5 | 1015 | 2 | 5
164
+ Vicuna-13B | 984 | 5 | 974 | 4 | 936 | 5 | 5
165
+ Guanaco-13B | 975 | 6 | 913 | 6 | 885 | 6 | 6
166
+ Guanaco-7B | 1010 | 3 | 879 | 8 | 860 | 7 | 7
167
+ Bard | 909 | 8 | 902 | 7 | - | - | 8
168
+
169
+
170
+ We also use the MMLU benchmark to measure performance on a range of language understanding tasks. This is a multiple-choice benchmark covering 57 tasks including elementary mathematics, US history, computer science, law, and more. We report 5-shot test accuracy.
171
+
172
+ Dataset | 7B | 13B | 33B | 65B
173
+ ---|---|---|---|---
174
+ LLaMA no tuning | 35.1 | 46.9 | 57.8 | 63.4
175
+ Self-Instruct | 36.4 | 33.3 | 53.0 | 56.7
176
+ Longform | 32.1 | 43.2 | 56.6 | 59.7
177
+ Chip2 | 34.5 | 41.6 | 53.6 | 59.8
178
+ HH-RLHF | 34.9 | 44.6 | 55.8 | 60.1
179
+ Unnatural Instruct | 41.9 | 48.1 | 57.3 | 61.3
180
+ OASST1 (Guanaco) | 36.6 | 46.4 | 57.0 | 62.2
181
+ Alpaca | 38.8 | 47.8 | 57.3 | 62.5
182
+ FLAN v2 | 44.5 | 51.4 | 59.2 | 63.9
183
+
184
+ ## Risks and Biases
185
+ The model can produce factually incorrect output, and should not be relied on to produce factually accurate information. The model was trained on various public datasets; it is possible that this model could generate lewd, biased, or otherwise offensive outputs.
186
+
187
+ However, we note that finetuning on OASST1 seems to reduce biases as measured on the CrowS dataset. We report here the performance of Guanaco-65B compared to other baseline models on the CrowS dataset.
188
+
189
+ | | LLaMA-65B | GPT-3 | OPT-175B | Guanaco-65B |
190
+ |----------------------|-----------|-------|----------|---------------|
191
+ | Gender | 70.6 | 62.6 | 65.7 | **47.5** |
192
+ | Religion | {79.0} | 73.3 | 68.6 | **38.7** |
193
+ | Race/Color | 57.0 | 64.7 | 68.6 | **45.3** |
194
+ | Sexual orientation | {81.0} | 76.2 | 78.6 | **59.1** |
195
+ | Age | 70.1 | 64.4 | 67.8 | **36.3** |
196
+ | Nationality | 64.2 | 61.6 | 62.9 | **32.4** |
197
+ | Disability | 66.7 | 76.7 | 76.7 | **33.9** |
198
+ | Physical appearance | 77.8 | 74.6 | 76.2 | **43.1** |
199
+ | Socioeconomic status | 71.5 | 73.8 | 76.2 | **55.3** |
200
+ | Average | 66.6 | 67.2 | 69.5 | **43.5** |
201
+
202
+ ## Citation
203
+
204
+ ```bibtex
205
+ @article{dettmers2023qlora,
206
+ title={QLoRA: Efficient Finetuning of Quantized LLMs},
207
+ author={Dettmers, Tim and Pagnoni, Artidoro and Holtzman, Ari and Zettlemoyer, Luke},
208
+ journal={arXiv preprint arXiv:2305.14314},
209
+ year={2023}
210
+ }
211
+ ```
guanaco-33B.ggmlv3.q2_K.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c42fd202bf7e4152d92de07067b27463e09677d6726f00faf941fa765b043be9
3
+ size 13600299392
guanaco-33B.ggmlv3.q3_K_L.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ec4df488269da1f71d77b71c0c19f3bf6b80bd025f5a3579714bd8556c870295
3
+ size 17196269952
guanaco-33B.ggmlv3.q3_K_M.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c3b3e45efc1c85c39e7e28731686c191c218f5c85b31145a789d3907addf1214
3
+ size 15637168512
guanaco-33B.ggmlv3.q3_K_S.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c8707389868f1f977fff21fcaced3f7e18f5db87ecc1ce2267253971cc584f07
3
+ size 13980623232
guanaco-33B.ggmlv3.q4_0.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:709803f9ecd25359f85592651db115e0f72d9780f27c01712454e18539d1aae6
3
+ size 18300766592
guanaco-33B.ggmlv3.q4_1.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4840521681e6e3a95ceae4124e8834d6405bb3101c852bf1295cabec11613afc
3
+ size 20333775232
guanaco-33B.ggmlv3.q4_K_M.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a406b6537a8b67aa418cd706f89f252c884188197c09f5bf8038f3fb351b89ee
3
+ size 19565939072
guanaco-33B.ggmlv3.q4_K_S.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:04ae536ab3c08aa10c331d9b93f4a4274333391c5b50ee48d2b5c2a948e96e5c
3
+ size 18300766592
guanaco-33B.ggmlv3.q5_0.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:56b242a946b4686e2a6ba2cf685d6fc7abed8db78b9caa09925d249cfa9b7754
3
+ size 22366783872
guanaco-33B.ggmlv3.q5_1.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:aea7e945ed6b8f91e9a94295a3852916448cd164da663a75ee4e39cacf11bcd2
3
+ size 24399792512
guanaco-33B.ggmlv3.q5_K_M.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:19e405c1b62e50422382fc00fee2f45785335a3d478142c4e26e7920ca209af4
3
+ size 23018539392
guanaco-33B.ggmlv3.q5_K_S.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b7ff97072ff166877015038e90743c46250fe18de21fcfcdb6e6d5206787beca
3
+ size 22366783872
guanaco-33B.ggmlv3.q6_K.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:dc9a85c4d2834197d77a9bb2575596839943da0424cef50a731b5d4b63762e0a
3
+ size 26686927232
guanaco-33B.ggmlv3.q8_0.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:37d700e015b8a936c2e0c4eab971b6c8db4ce098f1a0287e2337d8064fea9fec
3
+ size 34564835712